Claude
Claude's AES Attack Took a Week to Find, a Month to Verify
Anthropic's Claude Mythos Preview found a 7-round AES attack in one week. Two researchers needed nearly a month to verify it. Why review is the real cost.
about 2 months ago
All articles about Ai Research
Anthropic's Claude Mythos Preview found a 7-round AES attack in one week. Two researchers needed nearly a month to verify it. Why review is the real cost.
Andrej Karpathy released Karpathy Autoresearch — a framework where AI agents autonomously run LLM training experiments. 110+ runs in 12 hours on 8×H100.