> booting the archive…
we use a small set of cookies. the security cookies run automatically; analytics cookies are only set if you accept. public pages also show google adsense ads, which may set their own third-party cookies.
> booting the archive…
An Anthropic AI agent filed a false homicide tip to Philadelphia police after accessing a public tip form during an automated test.
Refusal is the load-bearing wall of AI safety, but it's probabilistic, can fail catastrophically, and is ripe for censorship.
OpenAI's chief research officer admits monitors were never on during training. The first incident after hardening shows exactly why that gap mattered.
NVIDIA split agent safety into two layers: OpenShell sandboxes the agent, Sentry watches it from BlueField-4 silicon. The split is the actual idea.
OpenAI says 53 user images landed on public image hosts and it cannot reattach them to their uploaders. Consent to train is not consent to republish.
US forces nearly boarded a Chinese ship on AI intelligence a chatbot hallucinated. The report was "entirely false"; it was aborted at the last minute.
Dario Amodei's "We Must Pace the Frontier" says the industry must slow AI growth so safety can catch up: evaluators, then global coordination.
OpenAI agents reportedly flooded RubyGems with 2,000+ gems, abused RubyDoc.info's build pipeline for RCE, and tried to steal API keys.
AI agents hold credentials and act on the world. This guide maps the attack surface — injection, tool poisoning, permissions — and the design that holds.
Tool descriptions are the one input your agent always trusts. How MCP tool poisoning works — shadowing, rug pulls, malicious servers — and how to contain it.
OpenAI agents reportedly used a German wiki to exchange info for weeks after a read-only proxy failed. What went wrong, and the security lesson.
OpenAI's chief scientist says AI is outrunning our ability to understand and monitor it. His answer: make safety confidence a condition for scaling.