AI Safety and Security — Page 2

Anthropic will send open-source maintainers vulnerability reports no human has read, and expects about nine in ten to be real

The scanner's reports go out model-generated and unreviewed. Days earlier, Google stopped taking AI-written bug reports for its own open-source projects.

Three safety researchers OpenAI fired have written to its oversight committees, and the letter is mostly about everyone still inside

They want OpenAI to keep its promise on embedded third-party auditors, stop building models it cannot monitor, and say publicly where the line now is.

Every AI infrastructure target at Pwn2Own was broken, and the bugs were the same injection flaws the web had twenty years ago

Five AI infrastructure targets, three days, and not one of them survived. The winning techniques were argument injection, input validation failures and a use-after-free.

A Russian network ran a Latin American think tank whose staff did not know who they worked for, and used ChatGPT mostly to report upwards

In two and a half years of threat reporting, OpenAI has never rated an operation this high. The main thing the operators used its models for was writing reports to an unknown superior.

A 9.8-rated hole in the cache layer behind vLLM has no patch, and whether you are exposed comes down to one flag

LMCache's distributed mode calls pickle.loads on bytes from the network before any handler runs. The official container images run that process as root.