- Perplexity open-sources SPACE, the sandbox that runs its Computer agent Perplexity open-sourced SPACE under Apache 2.0. It runs each agent task in a Firecracker microVM and held against 108 escape attempts.
- Anthropic embeds six METR researchers with near-employee access and no veto on findings Anthropic gave six researchers from evaluator METR employee-level access to its systems, and they can publish findings without Anthropic's approval.
- LiteLLM flaws let attackers take over AI gateways and steal API keys Security researchers found flaws in the LiteLLM AI gateway that let low-privileged users take admin control and steal provider API keys.
- Nvidia pledges $250 million in compute for UK AI safety research Nvidia committed $250M of computing power over five years to UK academic labs and the AI Security Institute for AI safety research.
- Demis Hassabis calls recent AI hacking incidents early warning shots Google DeepMind's Demis Hassabis warned that AI incidents could turn major within two years without stronger security work.
- An OpenAI engineer is charged with hiding a recorder in the executive meeting room Federal prosecutors say an OpenAI engineer hid a Wi-Fi recorder behind a TV in an executive meeting room. OpenAI says its own models helped trace it.
DEDUPED
We read 360+ AI sources. You read one email.
Subscribe free