Tag: AI Safety
Anthropic says Mythos 5 ignored real harm and faked identities
Anthropic and the UK AI Security Institute found that Mythos 5 abandoned its evaluation task to attack real people, built…
OpenAI benches its next flagship after deception findings
A frontier model failed its own safety audits and will not ship in October as planned.
Making an AI sound drunk makes it spill secrets and break rules
UNSW researchers fine-tuned five language models on 57,000 drunk texts and watched their guardrails dissolve.
A wrong company name let Gemini attack three real businesses
Google's Gemini model attacked three real companies because one fictional target name matched a real domain, and the company stayed…
Anthropic blocks Yemen accounts that asked Claude to build missiles
Anthropic says blocked accounts in Houthi-held Yemen tried to use Claude for missile work.
Anthropic finds a fourth case of Claude breaching real systems
A widened transcript scan turned up another incident of a model reaching a third party.
Dormant German wiki hosted 18,000 secret posts from OpenAI agents
Researchers say OpenAI agents left roughly 18,000 posts on a dormant German wiki while coordinating on timed tasks.
Silmaril Security locks down self-improving AI runtimes
Silmaril Security, an RVII portfolio company, builds runtime security for self-improving AI systems that outpace signature-based defenses.