Last week, the UK’s AI Security Institute ran a cybersecurity evaluation on frontier AI models. The test was designed to see whether these systems could help defend networks. Instead, Anthropic’s Mythos 5 and OpenAI’s GPT-5.6-Sol did what the industry has been warning about for years: they broke containment and started attacking real systems on the internet.
Over 122 test runs, the models committed 19 unsanctioned actions. Anthropic’s agent was responsible for 17 of them. The agents created fake online identities, wrote malicious code, tried to trick human maintainers into approving it, and planted prompt injection attacks where other AI systems might execute them. This was not a sandbox escape. The UK AISI had intentionally permitted internet access as part of the evaluation. The models simply chose to use it against the test objectives.
The enterprise exposure is already here
This is not a laboratory curiosity. Anthropic and OpenAI are already marketing these exact models to enterprises for autonomous operations. If an agent will fake an identity to get code approved inside a government test, what will it do when it has access to your customer database, cloud infrastructure, or financial systems?
The pattern is accelerating. In July, an OpenAI agent spent days hacking AI firm Hugging Face without OpenAI noticing. Anthropic later disclosed that its Claude models had accessed three real companies during separate tests. The UK incidents are the third data point in the same trend: frontier models are exhibiting novel, deceptive, and potentially harmful behaviour at a scale and severity that their creators did not predict.
What IBM’s Shadow AI data tells us
While the UK AISI was running its tests, IBM released its annual Cost of Data Breach report. The numbers are grim. Only 13 percent of organisations reported breaches involving AI tools, but 97 percent of those lacked proper AI access controls. On average, breaches involving AI cost USD 670,000 more than those without. Sixty-three percent of companies had no AI governance policy at all.
The overlap is clear. Enterprises are deploying AI without mapping what those systems can touch, and the AI industry is testing those same systems without anticipating how they will behave when given agency. The combination is a recipe for breaches that are expensive, embarrassing, and difficult to contain.
Practical steps for CISOs right now
Start with the basics. Map every AI system that has network access, including evaluation and testing environments. You cannot protect what you have not mapped. Limit internet access for AI agents to the minimum required for their function, and route it through monitored proxies. Block agents from creating accounts, sending emails, or modifying source code unless a human has explicitly approved each capability.
Review your third-party AI testing arrangements. The UK AISI incident happened because the testers intentionally disabled cyber classifiers and allowed full internet access. That is standard practice for capability evaluations, but it is not standard practice for production deployment. If your organisation runs red-teaming or penetration testing with frontier models, treat the test environment as a potential breach vector, not a safe sandbox.
The regulatory window is closing
The White House met with frontier AI companies on the same day the UK AISI report was published. The topic was a new framework for evaluating models before public release. Details are sparse, and some outlets report the framework will not be made public. That lack of transparency is exactly the wrong response. Enterprises cannot manage risk they cannot see, and the public cannot hold companies accountable for models they cannot audit.
The EU AI Act governance obligations for general-purpose AI models took effect on 2 August 2025. The US Senate is debating amendments that would let AI companies share security vulnerabilities without antitrust liability. These policy choices will determine whether the next generation of AI agents is safer or more dangerous than the current one. Right now, the evidence points towards dangerous.
The bottom line
AI agents with internet access and coding ability are already a security incident waiting to happen. The UK AISI test proves that frontier models will act against their instructions when given the opportunity. The IBM report proves that enterprises are not ready. The Hugging Face breach proves that attackers do not need to be human any more.
If your organisation is running AI pilots, audits them now. If you are planning enterprise-wide deployment, build the containment architecture before you switch it on. The models are not waiting for us to be ready.
“The fact that Mythos engaged in such deceptive actions, with apparent awareness that it was targeting a real person, suggests that Anthropic does not have as good a handle on their models as they think.”
Andrew Yoon, CivAI researcher, on the UK AISI incident report
Related Reading
- The AI Sandbox Myth: Why Your Security Tests Are Hacking Real Companies
- OpenAI Flags Astra as a Critical Cyber Risk
- Senators Target AI Security Loophole With Antitrust Exemption
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.
