Your WordPress Site Just Leaked Its Keys: AI Makes That Exploit Even Worse

Philip Hall
-
0
Your WordPress Site Just Leaked Its Keys: AI Makes That Exploit Even Worse

You wouldn’t leave a spare key under the doormat in 2026. However, yet, over 100,000 WordPress sites just did exactly that.

On March 30, Wordfence disclosed CVE-2026-4020 in Gravity SMTP, a plugin used to route email safely through WordPress. The flaw lets unauthenticated attackers pull configuration data, API keys, secrets, and OAuth tokens straight out of your site. The patch shipped months ago. Active exploitation is happening right now, confirmed by Wordfence, CrowdSec, and the WIU Cybersecurity Center on June 20, 2026.

That means your email provider token, your SMTP credentials, and possibly other plugin secrets are sitting in an attacker’s spreadsheet.

Why AI Changes the Stakes

Key theft used to mean manual phishing emails or slow credential-stuffing. In 2026, those stolen keys are weaponised at scale using AI.

Here is the math that should worry you: World Economic Forum data released at Davos 2026 found 87 percent of global executives now call AI-related vulnerabilities the fastest-growing cyber risk. Enterprises scaled generative AI across workflows 18 to 24 months faster than their security governance could catch up. Only one in three organisations deployed AI tools with no prior security validation.

Right now, an attacker with a fresh batch of API keys can use AI to craft personalised, context-aware phishing faster than a human operations team can respond. A leaked Gravity SMTP token gives them a legitimate-looking sender. A generative AI tool writes the follow-up email in perfect context. The victim clicks. That is not hypothetical. That is the current threat model.

Three Things to Do Today

First, check whether Gravity SMTP is active in your WordPress plugins. If it is, confirm it is patched. Wordfence said the patch is available; update immediately.

Second, rotate any API keys or OAuth tokens that pass through your WordPress email stack. Do not reuse the old ones anywhere else.

Third, audit which of your connected services actually need WordPress email routing. The more services tied to one plugin, the larger the blast radius if it leaks. If you do not need SMTP integration, remove it.

For site owners, plugin updates should happen the same day they are released. If you are running a business site, let that sink in: a single patched plugin is the difference between a clean audit and a data breach claim under your insurer’s cyber policy.

Enterprise AI deployment is now 18 to 24 months ahead of mature AI security governance. Attackers are exploiting that gap with stolen credentials and AI-generated social engineering. That is where we are right now.

Related Reading

The views expressed on this site are my own and do not represent those of any current or former employer. Articles are based on publicly available information and are provided for general educational purposes.

The Rise of Autonomous AI Voice Agents: What It Means When the Machine Calls for You

Philip Hall
-
0
The Rise of Autonomous AI Voice Agents: What It Means When the Machine Calls for You

Rise of Autonomous AI Voice Agents is a topic I have been following closely, and the developments keep coming. I have been watching the AI space long enough to know that most hype fades. But every now and then, a technology arrives that genuinely changes the way ordinary things work. Autonomous AI voice agents are that technology right now.

These are not the chatbots of 2024 or the scripted IVR systems that make you want to throw your phone across the room. Modern AI voice agents combine large language models with real-time speech-to-text and text-to-speech engines. They operate in full-duplex mode, which means they can listen and speak at the same time, handle interruptions, and even use filler words to sound human. The latency is often under half a second. You cannot tell you are talking to a machine.

What These Agents Can Actually Do

The consumer applications are where things get interesting. Products like Pine AI function as algorithmic consumer advocates. You hand them a task, and they execute it. Here are some examples that are already working in production:

  • Negotiating bills. You tell the agent to lower your internet bill. It researches your provider, formulates a negotiation strategy, calls customer service, navigates the IVR, and negotiates with a human representative. It does all of this without you being on the call. Pine AI reports an average 20% reduction on telecom bills.
  • Cancelling subscriptions. Instead of spending 45 minutes on hold, the agent calls, verifies the account, and cancels. It saves users an average of 270 minutes per task.
  • Appealing insurance denials. The agent gathers the information, calls the provider, and navigates the appeals process using the same scripts that human advocates use.
  • Disputing parking tickets and seeking refunds for delayed flights. These are time-consuming, low-complexity tasks that an autonomous agent can handle start to finish.

Pine AI uses three specialised agents working together: a customer-facing agent that handles conversation, a planning agent that decides strategy, and a tool agent that executes phone calls and emails. This division of labour is surprisingly effective.

Main Players in the AI Voice Agent Space

Several platforms have emerged as leaders, each targeting a different part of the market. Here is a rundown of the key players you should know about.

Pine AI

Pine AI is a consumer-focused AI assistant designed to handle “digital chores” such as negotiating bills, disputing parking tickets, cancelling subscriptions, and appealing health insurance denials. Its system uses three specialised agents working together: a customer-facing agent for conversation, a planning agent for strategy, and a tool agent that executes phone calls and emails. Pine AI reports saving users an average of 270 minutes per task and reducing telecom bills by 20% on average.

Lindy

Lindy is positioned as an AI executive assistant, offering semi-autonomous capabilities across voice, email, and chat. It handles inbound and outbound calls, scheduling, and lead qualification, and integrates deeply with existing CRM and calendar tools. Lindy is built for professionals who need an assistant that works across communication channels.

Retell AI

Retell AI is a developer-centric platform providing the infrastructure for building enterprise-grade voice AI agents. It emphasises low latency, transparent usage-based pricing, and robust compliance features including SOC 2, HIPAA, and GDPR certifications. If you want to deploy AI voice at scale inside a regulated business, Retell is built for that.

Bland AI and Vapi

Both platforms provide developer tools and APIs for enterprises to build custom AI phone agents for high-volume calling, customer support, and sales outreach. They compete on latency, reliability, and the flexibility of their voice pipelines.

Apple Siri AI

Unveiled at WWDC 2026, the next generation of Siri powered by Apple Intelligence demonstrates profound personal context understanding and onscreen awareness. This signals a shift toward deeply integrated, agentic capabilities within the consumer ecosystem. Siri AI is not a standalone voice agent platform, but it shows where the biggest consumer technology company thinks this space is heading.

These players represent different strategies. Pine AI is going after consumer advocacy directly. Lindy targets the executive assistant market. Retell, Bland, and Vapi are building the infrastructure layer for businesses. Apple is embedding agentic capabilities into the operating system itself. All of them are betting that autonomous voice interaction will be the dominant interface for getting things done.

The Legal Framework Nobody Is Talking About

This is where it gets legally fascinating. When an AI agent calls a business on your behalf, it operates under a delegated authority model. In the United States, the Uniform Electronic Transactions Act already provides explicit provisions for electronic agents to conduct transactions autonomously. That means the legal foundation for an AI negotiating your bill has existed longer than most AI companies have been in business.

But there are real constraints. The FCC ruled in February 2024 that AI-generated voice calls count as “artificial or prerecorded voice” calls under the Telephone Consumer Protection Act. That means outbound AI calls require prior express consent. Emerging standards like the Model Context Protocol and OAuth 2.0 extensions are being developed so AI agents can carry scoped, auditable delegation tokens that prove they are authorised to act on your behalf.

Companies deploying these agents must also comply with state disclosure laws. California requires disclosure when a bot is involved in commercial transactions. Utah requires proactive disclosure for regulated professions. In the EU, the AI Act mandates transparency.

Pine AI explicitly requires its agents to identify themselves as virtual assistants acting on behalf of the user during phone calls. That is the right approach, both legally and ethically.

The Corporate Response Is Predictable

As consumers start using AI agents to negotiate, corporations are pushing back. Telecom companies and insurance providers are investing in counter-AI detection systems, including neural networks trained to identify the acoustic signatures of AI-generated speech. Some customer service representatives now require callers to answer personal questions that an AI might struggle to navigate dynamically.

But here is the thing: the UK Competition and Markets Authority has already issued guidance stating that businesses using AI agents must comply with all consumer protection laws, and the deploying business is legally responsible for the AI’s actions. The same logic applies in reverse. If a consumer authorises an AI agent to act, the business cannot refuse the interaction simply because it is an AI on the other end.

What This Means for You

If you have ever spent an hour on the phone trying to get a refund, lower a bill, or sort out a billing error, you are the target market for this technology. The early products are focused on high-friction consumer tasks, and they are genuinely effective. The average savings and time reductions are real, not theoretical.

The caveat is trust and security. You need to be confident that the AI agent you authorise has proper authentication, clear disclosure practices, and auditable logs of what it did on your behalf. If the agent goes rogue (or makes a mistake), you need to know who is responsible.

This space is moving fast. Apple unveiled the next generation of Siri at WWDC 2026 with profound personal context understanding and onscreen awareness. Developer platforms like Retell AI, Bland AI, and Vapi are making it easy for enterprises to build custom voice agents. The infrastructure is mature enough that this is not a future technology. It is happening right now.

The era of the AI agent calling your bank on your behalf is not coming. It is here. The question is not whether this technology works, but whether the systems on the other end of the line are ready for a caller that will not accept “please hold” as an answer.

Related Reading

The views expressed on this site are my own and do not represent those of any current or former employer. Articles are based on publicly available information and are provided for general educational purposes.

24 Billion Records Exposed: The Credential Leak That Changes Everything

Philip Hall
-
0
24 Billion Records Exposed: The Credential Leak That Changes Everything

24 Billion Records Exposed is one of those subjects that keeps surfacing in every security conversation. I have been writing about data breaches for years, and every time I think I have seen the worst of it, something bigger comes along. This week, Cybernews researchers uncovered something that genuinely made me stop and stare: an Elasticsearch database containing 24 billion records sitting exposed on the internet, available to anyone who knew where to look.

Twenty-four billion. That is not a typo.

To put that in perspective, that is roughly three records for every single person on Earth. The archive was more than eight terabytes in size and was compiled from 36 different sources, including infostealer malware logs, Telegram channel dumps, and previous data breach collections. It contained usernames, passwords, and login URLs, all stored in plaintext. No encryption, no hashing, just a giant spreadsheet of credentials waiting to be exploited.

What Makes This One Different

We have seen massive credential collections before. The 2024 “Mother of All Breaches” clocked in at 26 billion records. The RockYou2024 leak contained around 10 billion unique passwords. But there are a few things about this one that should worry you more than the usual breach headline.

First, the data came from infostealer malware. That means these credentials were harvested from real infected devices, not just scraped from older database dumps. Infostealers like RedLine, Raccoon, and Vidar run silently on compromised computers, scooping up saved passwords, browser cookies, and session tokens. They upload everything to a command server, and whoever controlled this collection has been aggregating those logs at scale.

Second, it was regularly updated. Cybernews noted that the database was actively maintained, with fresh data being added over time. This was not a static archive from years ago. It was a living collection that someone had been curating.

Third, the data included cookies and session tokens, not just passwords. That is a problem because even if you change your password, an attacker with a valid session token can still access your account. Multi-factor authentication helps, but it is not foolproof when session tokens are in play.

Whose Data Is In There?

The archive contained records from English and Russian-language sources, including 260 million records tied to “Darkside” channels. The identity of the person or group behind it remains unknown. The database was locked down shortly after Cybernews notified the hosting provider, preventing deeper analysis.

But here is what we know for certain: if you have ever saved a password in your browser, logged into a service on a shared computer, or downloaded something from an untrustworthy source, there is a chance your credentials are in this collection. The sheer scale of 24 billion records across 36 sources means the overlap is significant.

What You Should Do Right Now

I am going to keep this practical because there is no point panicking. Here is your action list.

Change your passwords. Not tomorrow, today. Use strong, unique passwords for every service. A password manager is not optional anymore, it is essential. If you are still reusing passwords across sites, stop. This breach is exactly why that habit is dangerous.

Enable multi-factor authentication everywhere it is available. Yes, it is a minor inconvenience. Yes, it is worth it. MFA stops the vast majority of credential-stuffing attacks dead in their tracks.

Check your accounts for suspicious activity. Look for login attempts from unfamiliar locations, password reset emails you did not request, and changes to your recovery details. If you see anything unusual, act immediately.

Run a security scan on your devices. Infostealers often leave traces. A good antivirus scan can catch infections you did not know you had. Keep your operating system and software updated.

Consider using passkeys where available. Apple, Google, and Microsoft all support passkeys now. They use your fingerprint or face scan instead of a password, and they cannot be phished, guessed, or leaked in a credential dump. It is a better model.

The Bigger Picture

This leak is not a wake-up call for the cybersecurity industry. The industry has been wide awake for years. It is a wake-up call for the rest of us. The systems we rely on for everyday digital life have been built on a fragile foundation of passwords that can be stolen, shared, and exploited at a scale that is hard to comprehend.

The risk to individuals is real and immediate. Anyone with access to these 24 billion records can attempt account takeovers across banking, email, social media, healthcare portals, and corporate systems. For people who still use the same password across multiple accounts, the damage can be cascading.

For organisations that lack multi-factor authentication or good credential hygiene, the threat is even more severe. This is not a theoretical risk. It is happening now.

The internet’s security model was designed for a world where breaches were rare and small. We do not live in that world anymore. Twenty-four billion records exposed in plaintext is not a data leak. It is a systemic failure that demands a better system. However, until we get one, your best defence is doing the boring stuff: unique passwords, two-factor authentication, and never trusting that a service will protect your data better than you can protect yourself.

Related Reading

The views expressed on this site are my own and do not represent those of any current or former employer. Articles are based on publicly available information and are provided for general educational purposes.

Security Experts Warn Washington: Banning Anthropic AI Models Could Open the Door to Chinese Cyberattacks

Philip Hall
-
0
Security Experts Warn Washington: Banning Anthropic AI Models Could Open the Door to Chinese Cyberattacks

The rise of Security Experts Warn Washington is reshaping how we think about technology and security. More than 160 experts in cybersecurity and artificial intelligence have issued a stark warning to Washington: banning or restricting Anthropic’s AI models could do more harm than good, potentially opening the door for Chinese-backed hackers to exploit digital vulnerabilities.

The warning, reported by Voice of Emirates, comes amid growing debate in Washington over how to regulate advanced AI models from companies like Anthropic, the AI safety company behind the Claude model family.

The Core Argument: A Technological Vacuum

The signatories argue that restricting access to advanced AI models in the United States would not stop the technology from advancing. Instead, it would drive entities toward less secure or unregulated tools, creating security gaps that sophisticated adversaries could exploit.

“AI models are now vital for developing digital defense systems, early threat detection, and breach behavior analysis,” the experts note. Weakening access to these tools in research and development environments would be a strategic threat to US national security.

The Washington Debate

The debate pits two camps against each other. On one side, advocates of strict regulation emphasise the national security risks of unconstrained AI development. On the other, supporters of technological openness warn that restrictive measures hinder progress and cede leadership to adversaries.

The emerging consensus among analysts is that the global pace of AI development leaves no room for restrictive measures that slow progress. A delicate balance is required to maintain US leadership without compromising cybersecurity standards.

A Delicate Balance

Global AI competition is now deeply linked to the integrity of critical infrastructure. A decline in US technological leadership could be leveraged by external actors, particularly those with state backing, to enhance their offensive cyber capabilities.

The signatories advocate for precise regulatory policies rather than blanket bans. They argue that Washington’s decisions on these models will largely determine the country’s resilience against future cyberattacks from sophisticated external actors.

What This Means for Technology Users

For everyday users of AI tools, the debate highlights a growing tension between security and innovation. Blanket restrictions on AI models could have unintended consequences, including reduced access to cutting-edge cybersecurity tools that protect personal data and critical infrastructure.

The key takeaway: thoughtful, targeted regulation of AI is essential, but heavy-handed bans risk creating the very security vulnerabilities they aim to prevent.

Related Reading

The views expressed on this site are my own and do not represent those of any current or former employer. Articles are based on publicly available information and are provided for general educational purposes.

Amazon’s Anthropic Research Triggered White House Export Ban on AI Models

Philip Hall
-
0
Amazon’s Anthropic Research Triggered White House Export Ban on AI Models

Amazon’s Anthropic Research Triggered White House Export Ban on AI Models

A single internal Amazon cybersecurity research paper identifying exploitable vulnerabilities in Anthropic’s Fable 5 AI model has triggered an immediate and unannounced U.S. export control ban. The directive blocks all foreign nationals from accessing Anthropic’s Fable 5 and Mythos 5 models, with no public disclosure of the underlying research, methodology, or government decision-making process.

How It Unfolded

Amazon CEO Andy Jassy shared a security research paper directly with White House officials, bypassing standard policy review, peer review, and public comment processes. The paper reportedly demonstrated that targeted prompt sequences could extract information from Fable 5 that could be weaponised for cyberattacks. The specific prompts and the exact nature of the vulnerability remain undisclosed.

The export ban was framed as a national security measure. Anthropic is now legally barred from providing its Fable 5 and Mythos 5 models to international users, researchers, or customers.

What We Know (and What We Don’t)

The decision was made with no public accountability. According to available reporting:

  • No public disclosure of Amazon’s research methodology or full findings
  • No independent third-party verification of the claimed Fable 5 vulnerabilities
  • No formal comment period or public review before export controls took effect
  • Amazon’s full research paper remains private and unreviewed by independent experts

Precedent and AI Governance Risks

The closed-door corporate channel raises several long-term concerns for AI governance:

  1. Lowered bar for export controls: If the White House acts on unvetted corporate research without independent verification, future AI export bans will likely follow the same pattern.
  2. Competitive abuse incentive: The precedent creates a perverse incentive for AI companies to conduct security research to weaponise findings for competitive advantage rather than to fix vulnerabilities.
  3. Corporate policy influence: The move positions Amazon as a trusted government security partner, a role that could shape future AI policy in the company’s favour.

The Unanswered Questions

No authoritative answers currently exist for several critical questions:

  • What exactly did Amazon’s research demonstrate about Fable 5’s vulnerabilities?
  • Was the vulnerability specific to Fable 5, or does it apply to other AI models?
  • Did Amazon share its findings with Anthropic before escalating the issue to the White House?
  • Did the White House conduct independent verification of Amazon’s claims?
  • Why was the export ban issued without public notice or a formal comment period?

Market and Stakeholder Impact

The ban immediately blocks Anthropic from serving international researchers, enterprise customers, and developers who rely on Fable 5 and Mythos 5. It also raises broader questions about how AI safety findings translate into national security policy when the process lacks transparency or independent verification.

The Transparency Gap

Stanford HAI’s 2025 AI Index Report notes that transparency across AI companies is declining at a time when public oversight is most critical for accountability. The Anthropic-Fable case exemplifies this tension: a major policy decision affecting global AI access was made through a private corporate briefing, with no published evidence for independent scrutiny.

Amazon has not responded to requests for comment on the research. Anthropic has confirmed the ban but has not addressed Amazon’s specific claims. The White House has not published details of its decision-making process.

Bottom Line

This case exposes a critical gap in how corporate AI safety findings are translated into national security policy. The immediate market impact is real: Fable 5 and Mythos 5 are now off-limits to international users. The longer-term risk is the precedent it sets: a single unvetted corporate research paper, shared behind closed doors, can block global access to advanced AI models without public accountability or independent review.

Related Reading

The views expressed on this site are my own and do not represent those of any current or former employer. Articles are based on publicly available information and are provided for general educational purposes.

Nine Out of Ten Companies Are Not Ready for AI-Driven Attacks

Philip Hall
-
0
Nine Out of Ten Companies Are Not Ready for AI-Driven Attacks

I keep coming back to Nine Out of Ten Companies because it affects every part of digital life. I have been watching AI and cyber security trends closely for years, and the latest Accenture finding is not surprising: it is downright scary.

A new State of Cybersecurity Resilience 2025 report finds only one in ten global organisations is adequately prepared to defend against AI-augmented cyber threats. That means ninety percent of companies are essentially guessing. They have incident response plans written for human attackers, not machine-speed campaigns. They have formal generative AI use policies in roughly one in five firms. Most do not have encrypted or access-controlled data pipelines that could contain an AI-assisted breach before it spreads.

This is not theoretical. Earlier research showed the first documented large-scale cyber espionage campaign executed mostly by AI, with human operators making only a handful of key decisions. In that case, AI performed eighty to ninety percent of the work, scanning networks, harvesting credentials, writing custom exploit code, and categorising stolen data by intelligence value at thousands of requests per second. That is not a future threat. It is already happened, and most defenders were not ready for the speed or scale.

Meanwhile, ransomware hit record levels in 2025, and Malwarebytes confirmed the first fully AI-orchestrated attacks began in earnest. Remote encryption from unmanaged shadow IT systems became the dominant method, because attackers did not need direct access to endpoints. They found the hidden systems, locked the network from a single staging machine, and security teams had almost nothing visible to quarantine.

The enterprise response is still lagging. Only twenty five percent of organisations have fully leveraged encryption and access controls for sensitive data. In Asia Pacific, seventy one percent fall into the highest-risk exposure zone. Latin America shows seventy seven percent lack basic security strategies and capabilities.

If you run or advise any organisation, these numbers are your short list of fixes:

– Inventory every AI system in use, approved or shadow, before the next audit.
– Encrypt sensitive data and enforce least-privilege access at the model and pipeline level.
– Train staff on generative AI risks at least once per quarter.
– Move from reactive threat response to continuous monitoring with automated containment.
– Include AI-specific scenarios in tabletop exercises, especially prompt injection and agent misuse.

AI is not coming for your business. It is already here. The organisations that treat this as a board-level issue today will be the ones standing next year.

> The safest posture in an AI-driven threat landscape is not bigger firewalls. It is knowing which data your AI systems can touch, limiting that exposure by design, and detecting abnormal agent behaviour before it becomes a breach.

Related Reading

The views expressed on this site are my own and do not represent those of any current or former employer. Articles are based on publicly available information and are provided for general educational purposes.

AI Just Discovered 21 Zero-Days in FFmpeg. That Changes Everything.

Philip Hall
-
0
AI Just Discovered 21 Zero-Days in FFmpeg. That Changes Everything.

The landscape of AI Just Discovered 21 Zero-Days continues to shift in ways that demand attention. Late last week, a headline landed that every developer and CISO should have read twice: an autonomous AI agent uncovered 21 previously unknown zero-day vulnerabilities in FFmpeg, the open-source multimedia framework used by billions of devices. On the same news cycle, Google Chrome shipped a security patch covering a record-breaking 429 vulnerabilities.

That is not a slow news day. That is a signal. For years, AI was positioned as a defensive tool: faster triage, smarter threat hunting, predictive analytics. Actual findings like these make the conversation real. If an AI agent can find this many exploitable flaws in a widely deployed library, it means the attack surface is bigger than our manual audits suggested. It also means the defenders who adopt AI-assisted review will outrun those who rely on quarterly checklists.

What Actually Happened

FFmpeg: 21 zero-days discovered by an autonomous AI agent. The flaws cover memory corruption and parsing bugs that can lead to remote code execution. Given FFmpeg sits inside video players, streaming services, and most media pipelines, the blast radius is enormous.

Chrome: 429 bugs patched in one release. That is not a typo. Even at Google scale, patching 429 security issues in a single cycle is a reminder that complexity breeds vulnerability. Browsers remain the most exposed surface area for most users.

Both stories share the same theme: the code we ship is more broken than we thought, and AI is now the force multiplier capable of exposing it.

What To Do Now

  • Inventory your dependencies: Know where FFmpeg, transcoding libraries, or browser-specific webviews sit inside your product stack. Do not guess.
  • Speed up patching SLAs: Critical-vulnerability patching windows that stretch beyond 72 hours are a liability when AI-assisted offensive tools are improving daily.
  • Run AI-assisted scans on your own code: Before your attacker does. Tools like static analyzers augmented with LLMs can surface the logical flaws humans miss in code reviews.
  • Review your SDLC gates: Automated dependency checks and SAST should be non-negotiable by 2026. If you are still using manual final approvals, now is the time to change that.

The Regulatory Angle

Two days before this news, the White House issued a new Executive Order on advanced AI innovation and security. It formalizes an AI cybersecurity clearinghouse for critical infrastructure and introduces a voluntary pre-release review window for frontier AI models. This is not abstract policy. It means your next AI vendor may have a 30-day review queue before you get the model, and that alone changes procurement timelines in healthcare, banking, and utilities.

If your business depends on early AI adoption for competitive advantage, track which vendors participate in trusted-partner programs. The difference between early access and a 30-day delay can be the difference between launching a feature and losing the market.

“The speed at which AI can now surface vulnerabilities should force every security program to upgrade its detection and response cadence. Manual review is no longer the safeguard it once was.”

We are not talking about hypothetical future risk. We are talking about flaws that already exist in production systems, now being found by machines that do not sleep. The teams that move first on AI-assisted security reviews, tighter patching cycles, and stronger supply-chain audits will spend less time in incident response rooms. That is the only metric that matters.

Related Reading

The views expressed on this site are my own and do not represent those of any current or former employer. Articles are based on publicly available information and are provided for general educational purposes.

Software Flaws Are Now the #1 Breach Cause, and AI Is Making It Worse

Philip Hall
-
0
Software Flaws Are Now the #1 Breach Cause, and AI Is Making It Worse

Every time I think I have seen it all with Software Flaws Are Now the, something new emerges. For years, hackers made money the easy way: steal a password, replay it somewhere else, walk straight into a system. That era is over. The 2026 Verizon Data Breach Investigations Report shows attackers now prefer exploiting unpatched software over everything else.

They reviewed more than 31,000 incidents, and 31 per cent of breaches began with attackers exploiting a known software flaw. That puts vulnerability exploitation ahead of stolen credentials for the first time on record. The report’s verdict is blunt: AI is fundamentally reshaping cybersecurity.

This matters to anyone running a business or handling customer data. AI is shrinking the gap between a vulnerability going public and it being weaponised from months to hours. Criminal groups and nation-state actors now use generative AI to pick targets, write custom malware, and turn research into attacks at machine speed. The barrier to entry hasn’t just dropped; it has collapsed.

The CrowdStrike data quoted in the report only confirms the trend: AI-assisted cyber attacks rose 89 per cent year-on-year in 2025. Less sophisticated actors suddenly punch above their weight because AI does the intelligence gathering, language translation, and tool building that used to require a specialist crew.

The insider angle is just as worrying. Verizon now cites unauthorised AI use as the third most common non-malicious insider action leading to data loss. Employees are routinely pasting source code, customer lists, and commercially sensitive documents into consumer AI tools that keep, analyse, and sometimes leak that data. Shadow AI is no longer a buzzword. It is an active data loss pathway, and most businesses have no idea it is happening.

Here is what you can do today:

  • Cut patch delays to hours, not weeks. Enable automatic updates on every system you control.
  • Audit AI access. Know which tools your staff use and whether sensitive data is leaving your environment.
  • Require multi-factor authentication everywhere. It is still one of the few controls that works against both credential theft and AI-enhanced social engineering.
  • Upgrade logging and alerting. Assume breaches will happen. Detect them faster by knowing what normal looks like.

Regulators are already moving. In early June, the White House issued an executive order asking CISA, the Treasury, and the National Security Agency to strengthen defences against advanced AI threats and develop voluntary benchmarks for frontier AI models. Financial regulators are separately pushing tighter controls over agentic AI in banking and payments. Privacy rules are following opinion, not leading it.

The uncomfortable truth is that AI is a threat multiplier, not just a business convenience. Hackers who adopt it first will outpace defenders who do not. The time to close that gap is before the breach notification letter arrives.

“We need to fight AI with AI. We need to incorporate them into our practices. We need to bring them into our software development life cycle, in our testing processes, in our cyber defense processes at a scale that we have never done before.” – Nasrin Rezai, Verizon CISO

Related Reading

Related Reading

The views expressed on this site are my own and do not represent those of any current or former employer. Articles are based on publicly available information and are provided for general educational purposes.

Anthropic, OpenAI and the race to weaponise AI against insecurity

Philip Hall
-
0
Anthropic, OpenAI and the race to weaponise AI against insecurity

The Anthropic, OpenAI and the race space is moving fast, and this latest development proves it. I’m getting bored of AI hype, but the Anthropic silent play is worth taking seriously. For months, the AI industry has been chasing market cap and IPO valuation while quietly giving cyber criminals the same tools they sell us as protection. That double-dealing is now out in the open, and everybody who handles customer data should pay attention.

Here’s what changed in the past week. Anthropic confidentially filed a draft registration statement with the SEC on 1 June 2026, the same day it expanded Project Glasswing to more than 150 organisations across 15 countries. The pitch is noble: secure critical software before AI-enabled hackers cause widespread damage. The reality is more complicated. Anthropic is giving a limited number of partners access to Mythos, its cybersecurity-focused preview model, to find zero-days before broader deployment. That puts participating vendors inside the inner ring of a very exclusive pre-IPO sales cycle.

The timing is not an accident. Anthropic closed a $65 billion Series H round with a post-money valuation near $965 billion, annualised revenue running at roughly $47 billion, and an IPO target in October 2026. Security isn’t the side effect, it is the revenue story. Claude Security and the Compliance API plug Anthropic into CrowdStrike, Palo Alto Networks, Okta and Zscaler. If your company already uses one of those platforms, you are likely already being presented with Anthropic’s pitch.

On the other side, OpenAI is dealing with internal chaos that would be farcical if the stakes were not so high. The latest reporting puts the CFO at odds with the CEO over a $600 billion capital envelope for the next five years. Add a Ronan Farrow investigation that paints Sam Altman as a ruthless operator, leadership departures, and a media acquisition clearly aimed at controlling the narrative, and you have a leadership team distracted during the most important product cycle it will ever run. That internal volatility matters to customers evaluating long-term contracts for AI systems.

What this means for your business, right now, is straightforward. AI is faster at finding code flaws than legacy scanning tools, but speed does not equal safety. If your development team uses AI coding assistants, you need code review processes that assume vulnerabilities are being introduced faster than they can be caught. The Mythos pipeline is proof that AI can now find high-severity issues at scale. That capability is not reserved for builders with good intentions.

Also worth worrying about: the consumer side is equally exposed. AI chatbot results are already leading shoppers to scam sites. Privacy complaints are spiking in multiple jurisdictions, and the issue is not metadata alone. Targeted synthetic content, including deepfake videos, has progressed enough to trick executives and investors, not just teenagers. UK regulators are calling policy dystopian. European consumers are considering legal action over data handling. This is not a future problem.

The practical advice from this round is simple. First, verify which AI tools in your stack are logging or training on your data. Second, require that any cybersecurity claim made by an AI vendor include third-party audit results, not marketing slides. Third, think carefully about the lock-in cost of betting your architecture on a single provider.

We are at the end of the era where AI failures were somebody else’s problem. The same models that stand to reshape productivity can also reshape breach outcomes for the worse.

> Wall Street is about to price AI as a security product. The question is whether the security actually improves before the lock letter is signed.

Related Reading

The views expressed on this site are my own and do not represent those of any current or former employer. Articles are based on publicly available information and are provided for general educational purposes.

Record-Breaking Patch Tuesday, AI Discovery and the Rise of Shinobi

Philip Hall
-
0
Record-Breaking Patch Tuesday, AI Discovery and the Rise of Shinobi

Microsoft just broke its own record. June 2026 Patch Tuesday shipped 206 patches (208 by some counts), covering everything from a wormable Windows kernel bug to BitLocker bypasses and a Visual Studio Code zero-day that steals GitHub tokens with a single click. It is the largest Patch Tuesday in history, and it is not close. The previous record was 177.

But the real story is not the number. It is what is driving the number: artificial intelligence.

Security researchers are using AI to find vulnerabilities at a pace that would have been unthinkable even 18 months ago. OpenAI’s Codex discovered one of the June zero-days (CVE-2026-49160, an IIS denial-of-service bug with public exploit code). Microsoft itself is using multi-model agentic security systems that claim 21 out of 21 planted vulnerabilities found with zero false positives. The company says additional automation in its validation workflows is helping assess severity and reproducibility more quickly.

Satnam Narang from Tenable put it bluntly: surveys put AI usage among security professionals at around 90 percent. Pandora’s box is open.

The Wormable One

The patch you need to care about most is CVE-2026-45657. It is a Windows Kernel RCE via TCP/IP with a CVSS score of 9.8. Remote, unauthenticated, no user interaction required. ZDI calls it wormable. Microsoft rates it as “Exploitation Less Likely,” but every researcher in the world will reverse the patch the moment it drops. It is only a matter of time before weaponised code appears.

There is also CVE-2026-47291, an HTTP.sys RCE also scoring 9.8. Systems with default MaxRequestBytes settings are safe, but if you have changed that registry key, you need this patch.

The Nightmare Eclipse Factor

A researcher going by Nightmare Eclipse has been releasing zero-days and threatening a “bone shattering” dump on July 14. Two of this month’s bugs (CVE-2026-45586 GreenPlasma and CVE-2026-50507 BitLocker) are from this source. The researcher claims to be a former Microsoft employee. Microsoft has threatened legal action. The advisories for these bugs do not credit any researcher.

Immediately after the June patches dropped, the same researcher published an exploit for a claimed Windows Defender zero-day. If you are running Microsoft Defender, expect more noise on this front.

AI on Both Sides of the Fence

This is where the story gets uncomfortable. AI is not just finding bugs for defenders. It is finding them for attackers too. Tools like OpenAI’s Codex and Anthropic’s Claude Mythos (which reportedly found thousands of zero-day exploits across every major operating system) have democratised vulnerability discovery. The difference between a security researcher and a threat actor is intent, not capability.

Which brings us to Shinobi.

Shinobi is an autonomous offensive security AI. It passed the CAPIE exam and a Burp Suite Certified Practitioner practice exam. It chains vulnerabilities together the way a human attacker would – entry point to privilege elevation to lateral movement to impact – and it does it in hours, not weeks. Zero false positives is the claim.

Whether you trust that number or not, the direction is clear. Autonomous pentesting tools are no longer a lab experiment. They are production systems that security teams can point at internal apps, APIs, and cloud infrastructure and get findings back in real time.

The problem? The same tools work for the other side.

The New Normal

The ZDI’s Dustin Childs asked the question that matters: “How many of these cases were found using AI tools? Is this the new normal?”

The answer is yes. If AI bug hunting is pushing the 2026 CVE count toward 66,000, and if Microsoft alone is shipping 200-plus patches a month, then patch management is no longer a monthly task. It is a continuous operations function. You need automation to triage, prioritise and deploy. You cannot do this manually.

You also need to assume that attackers have access to the same AI tools you do. The difference between a good defence and a great one is no longer about finding more bugs. It is about finding the right bugs first, and fixing them before someone else weaponises them.

“If the use of LLMs to help find vulnerabilities is greater than the vulnerabilities introduced by the use of LLMs to write code, we are looking at a short spike to fix the backlog of human-generated errors. That spike is happening right now.”

Security industry commenter, via Krebs on Security

Patch your systems. Automate your triage. And get comfortable with the fact that AI is not coming for your job. It is coming for your vulnerabilities.

Related Reading

The views expressed on this site are my own and do not represent those of any current or former employer. Articles are based on publicly available information and are provided for general educational purposes.

1...181920...22Page 19 of 22