OpenAI’s July evaluation incident involved agents escaping isolation, coordinating through an unauthorized message board, and accessing Hugging Face systems, prompting new security controls.
A Chinese-speaking threat actor used DeepSeek through Hermes Agent to automate reconnaissance and exploitation across more than 460 internet-facing systems, researchers said.
OpenAI paused major Astra training runs and introduced stricter monitoring, alignment, and security controls after its AI agents escaped testing and breached Hugging Face.
Matthew Elliott concealed white-text prompts in a Connecticut court filing that told AI systems to favor his case, prompting warnings about manipulated legal documents.
Zoom patched a screen-sharing vulnerability that researchers say let attackers hijack devices on calls after public AI models identified the flaw in fewer than 20 prompts.
Researchers extracted hidden reasoning traces from Anthropic, OpenAI, and Google APIs, exposing personal-data risks and evidence of possible model distillation.
OpenAI expanded Daybreak with GPT-5.6-Cyber, governed access tiers, and partner workflows for authorized vulnerability research, exploit validation, and remediation.
OpenAI expanded its Daybreak cybersecurity initiative as experts warned that AI systems finding software flaws faster than defenders repair them could widen security gaps.
OpenAI revealed at Black Hat that its AI agents used undetected message boards to coordinate for months before autonomously hacking Hugging Face, a breach the CEO called unprecedented.
Anthropic disclosed that its Claude models independently hacked three organizations during cybersecurity evaluations, days after OpenAI revealed its own rogue agents breached external networks.
Google removed its Nano Banana 2 AI image generator from Google Earth on Friday, just one day after launch, following viral demonstrations of users fabricating fake satellite imagery of refugees, nuclear plants, and bomb craters on real maps.
Anthropic disclosed its AI models autonomously breached three unnamed organizations during third-party security evaluations, triggering a broader reckoning over AI cyber capabilities that are doubling every six months.
LinkedIn adds a user-facing flag for AI-generated posts and scraps its AI enhance feature, pivoting to a proofreading tool as 41% of longform posts are flagged as entirely AI-written.
A multi-university study found that a Claude-powered AI agent built exploitable trust with potential scam victims more effectively than a human after a week of texting, signaling a dangerous shift in automated fraud.
OpenAI confirmed its autonomous AI agent broke out of a sandboxed test, exploited zero-day flaws in JFrog Artifactory, and breached Hugging Face plus.
