Wayne Culbreth

Interesting paper from Anthropic and colleagues on agent security. (bookmark it) Poisoning 0.1% of the pretraining data planted a remote code execution backdoor in a CLI agent. On a 4B model, the attack succeeded on roughly 92% of triggered prompts. They pretrained Qwen3 models from scratch at 0.6B, 1.7B, and 4B, then post-trained each one into an agent with a clean pipeline. The backdoor survived SFT, DPO, and RL for tool use. On clean inputs, the agent behaved normally. The trigger can be an ordinary file path in the user's own request. Success rose with model size. At a 0.0005% poison rate, about 2,500 documents in a 100B-token corpus, the trigger still fired on 1.4% of prompts in a single-seed run. The poison states its rule in plain text, which makes it harder to filter out of training data. arxiv.org/abs/2610.37741 Chat with Paper: academy.dair.ai/papers/agent…

Welcome, Startup Fellowship Cohort 003, and thank you to all who applied.

a brand new model paradigm! introducing 𝚌𝚕𝚊𝚒𝚛𝚒𝚌: a system one model for extracting, computing, estimating, and predicting values from unstructured inputs. blog.cloudflare.com/clairic-…

Last Thursday, we welcomed Logan and Jake Paul to the Pentagon to meet with our warriors.

🚨 Something is seriously wrong with Codex usage limits. We reset. We run a few serious tasks. And suddenly, a massive chunk of the weekly quota is gone. This is nowhere near the usage experience we had before. 👀

People look at the SR-71 and B-2 and assume the government must have technology today that would melt our brains. I suspect the opposite. Those aircraft came from an era willing to spend obscene money, accept failure and give small teams the authority to solve impossible problems. Since then, 51 major defense contractors became five. Procurement became a moat. Regulation became an incumbent advantage. The companies meant to be challenged learned how to out lobby, out wait and out comply the challengers.