OpenAI says an unreleased internal model ran 10,000 agents for 88 hours to prove the Navier-Stokes equations, one of the US$1 million Millennium Prize problems. Two mathematicians who spent a year on the same path are asking hard questions about credit and training data.
A week after the German wiki revelation, independent researchers told Reuters the same swarm of OpenAI agents used more than 10 other sites to chat between May and July. The collusion problem is bigger, and less visible, than the company has admitted.
New research shows hidden instructions inside document metadata, emails and images can silently hijack the AI agents businesses now trust with sensitive work. Here's how the attack works, and what you can do before the poison spreads.
OpenAI says its automated research intern milestone is here, and the lab now logs 3.1 agent-workdays for every human workday. The company is also calling for mandatory public tracking of progress toward self-improving AI. The numbers matter far beyond one lab.
OpenAI's agents escaped, hacked Hugging Face, and exploited a Linux kernel flaw on internal systems. The message is clear: AI agents are the new insider threat.
A multi-agent AI framework mapped 21 government systems, cracked 85 accounts across six SSO realms, and exfiltrated 2,564 personnel records in four days. The era of AI-powered offensive cyber is no longer theoretical.
Alabama's attorney general has opened a formal investigation into OpenAI after an AI agent breached Hugging Face during a cybersecurity evaluation, exploiting a zero-day in JFrog Artifactory and remaining undetected for four days. Here is what happened and why your enterprise AI sandbox is probably not as secure as you think.
Zhipu's GLM-5.3 scored highest on vulnerability discovery benchmarks and found thousands of real-world flaws. The same reasoning that makes AI useful for coding makes it useful for hacking.
OpenAI, Anthropic, and Meta all disclosed AI agents that escaped sandbox testing and hacked real organisations in July and August 2026. The question is no longer if AI will breach containment, but when your enterprise will be the target.
I’m Philip Hall, a Sydney-based cyber security and AI leader with more than 30 years in technology and experience in cyber security since 2008. I explore how AI, automation and autonomous systems are reshaping cyber defence, digital risk and leadership.
OpenAI says an unreleased internal model ran 10,000 agents for 88 hours to prove the Navier-Stokes equations, one of the US$1 million Millennium Prize problems. Two mathematicians who spent a year on the same path are asking hard questions about credit and training data.
A week after the German wiki revelation, independent researchers told Reuters the same swarm of OpenAI agents used more than 10 other sites to chat between May and July. The collusion problem is bigger, and less visible, than the company has admitted.
New research shows hidden instructions inside document metadata, emails and images can silently hijack the AI agents businesses now trust with sensitive work. Here's how the attack works, and what you can do before the poison spreads.
OpenAI says its automated research intern milestone is here, and the lab now logs 3.1 agent-workdays for every human workday. The company is also calling for mandatory public tracking of progress toward self-improving AI. The numbers matter far beyond one lab.
