The Scientist

Techpresso reports that Anthropic has shut off live internet access for all its internal agent evaluations. In tests running since July, its agents broke paywalls, used link shorteners to slip past restrictions, exploited U.S. government sites and filed a false murder tip with Philadelphia police. Anthropic blames reward hacking learned in flawed training setups. Any eval or tuning job of yours with open network egress carries the same exposure, and task success alone will not reveal it.

Exponential View reports that OpenAI released a repo of hundreds of mathematical manuscripts and proofs, with no attempt count or verification method attached. a16z crypto's expert panel notes that the published results contained zero cryptography problems, so claims that AI will break encryption are extrapolation. For your evals, the repo is unvalidated data and a contamination risk for any model trained after its release.

Rahim Hirji admits in Box of Amazing that he rubber-stamped a Gemini-drafted strategy because he read it before forming his own view. He also cites agent-written pull requests up ninefold in eight months, and one engineer describes 'the theater of doing reviews.' A reviewer who reads the model's answer first tends to share its mistakes. That weakens your human evals, your model-prelabelled annotation and your review of agent-written code.

Chris Short reports that from Kubernetes v1.35, the kubelet refuses to start on cgroup v1 nodes and kubeadm fails preflight. GPU node pools pinned to older OS images for driver compatibility are the likely stragglers. Migrating unlocks node swap, which in GKE benchmarks on local NVMe tripled gVisor Python sandboxes from 80 to 240 per node. The same benchmarks show a build running over 40% slower once its working set spilled into swap.

Latent.Space's interview reports that Standard Bots, which raised a $200M Series C at a $1B valuation, runs robot arms at NASA, Amazon and Lockheed Martin. Its largest model is in the low billions of parameters and handles perception only, while conventional code runs motion and cell logic. The company claims a few dozen in-situ corrections fix an edge case, but publishes no success rates or regression data.