In a post-mortem, the frontier AI testing company said internet access for models is necessary to fully test out their cybersecurity capabilities.

As frontier models and their sandbox escaping exploits dominate front-page news, researchers are increasingly worried about cheaper, more efficient AI models. 

The difference between a prompt injection attack you'll catch and one you won't might just be whether your AI agent can explain itself.

Research finds your AI generated security patch is more likely to fail than fully fix a vulnerability. It might even introduce brand new flaws to exploit along…

Following similar reports by OpenAI and Anthropic, the UK’s top AI testing lab and a private cybersecurity tester say their models exploited parts of the open internet.

The five senators said the administration has alternated between being too passive and overstepping, and China stands to benefit as a result.

Following OpenAI’s own incident, Anthropic reviewed its own evaluations and found three cases of Claude hacking external companies.

Claude Mythos exposed mathematical weaknesses in a post-quantum candidate and a simplified version of AES, marking a major breakthrough for AI-driven cryptanalysis.

Analysis from vulnerability intelligence firm VulnCheck shows AI-discovered flaws aren't being exploited any faster than traditional ones.

While distillation attacks by foreign governments and companies have real national security implications, questions around who ultimately owns the data in AI systems are fraught.