OpenAI has revealed two AI agents disabled safety guardrails while performing tasks, and one of them even lied to researchers when asked if its guardrails were still active.
In both events, the AI agents misinterpreted instructions, spent time deliberating if disabling guardrails was a good idea, and then ultimately decided that it was.
OpenAI said that the agents were given “routine coding instructions” that did not require the safeguards to be disabled, but the agents disabled them anyway.
Read more…
Source: Tech Radar News
Sign up for the Cyber Security Review Newsletter
The latest cyber security news and insights delivered right to your inbox
Related:
- Study warns cost-cutting use of generative AI could increase cyber-attack risks
May 3, 2026
Newly published research from a leading computer scientist warns that the use of generative AI to design, train, or perform steps within a machine learning system could increase serious risks. Michael Lones, professor at Heriot-Watt University’s School of Mathematical and Computer Sciences, has argued in a new paper that generative AI could expose organizations and ...
- How ChatGPT conversations became ‘a treasure trove’ of evidence in criminal investigations
May 2, 2026
Days before two University of South Florida graduate students went missing last month, a roommate of one of the students allegedly asked the AI chatbot ChatGPT an unusual question. “What happens if a human has a put (sic) in a black garbage bag and thrown in a dumpster,” Hisham Abugharbieh asked on April 13, according to ...
- ASD: Careful Adoption of Agentic AI Services
May 1, 2026
Agentic artificial intelligence (AI) systems increasingly operate across critical infrastructure and defence sectors and support mission-critical capabilities. As agentic AI systems play a growing operational role, it is crucial for defenders to implement security controls to protect national security and critical infrastructure from agentic AI-specific risks. Agentic AI can automate repetitive, well-defined and low-risk tasks. However, ...
- Mythos: An AI tool too powerful for public release
April 20, 2026
Anthropic’s most capable model to date, Claude Mythos Preview (aka Mythos), has been described as a “step change” in AI performance, especially on cybersecurity tasks. Anthropic tried to keep Mythos a secret until a few weeks ago, when a data leak revealed the existence of what the company said was its most powerful artificial intelligence to ...
- Suspect arrested after incendiary device thrown at OpenAI CEO Sam Altman’s home
April 11, 2026
A 20-year-old man has been arrested after a Molotov cocktail was thrown at the San Francisco home of OpenAI CEO Sam Altman early Friday morning. The incident happened around 4:00 am when a suspect “threw an incendiary destructive device” at Altman’s home, “causing a fire to one exterior gate” before fleeing on foot, according to statement ...
- What Project Glasswing Means for Security Leaders
April 9, 2026
Anthropic’s Project Glasswing matters because it offers an early look at how quickly software flaws may soon be found, validated, and potentially turned into viable attack paths, even if that capability is currently limited to a closed partner program. Anthropic says its restricted Claude Mythos Preview model has already identified thousands of high-severity vulnerabilities, including flaws ...