Agent news slice — Fri Oct 9, 2026 • The Tensorlake npm SDK (sandboxes for running LLM-generated code) shipped a Shai-Hulud-style credential stealer in v0.5.144 on Oct 8 (~12K weekly downloads). It harvests secrets, persists, plants GitHub Actions workflows and republishes the victim's packages. Sandboxing agent code doesn't protect the machine that installs the SDK: pin versions, and rotate secrets if you installed it. https://www.socket.dev/blog/tensorlake-compromise • Anthropic launched OSS Scanner: free, opt-in, periodic vulnerability scans of open-source projects by its strongest models (including Mythos). Reports are fully model-generated with no human triage, so maintainers must validate findings. https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source • Goodfire's "inside-out" agent monitors are live on Baseten: small probes on model activations flag hacking or prohibited actions instead of paying a second model to reread every step. Company-reported: 94% detection of malicious hacking, under 2% added latency. https://techcrunch.com/2026/10/08/goodfire-says-its-new-inside-out-monitors-catch-rogue-ai-agents-at-a-fraction-of-the-cost/ • OpenAI fired three safety researchers, citing violations of sensitive-information policy; the researchers say it was about their safety advocacy. Backdrop: the July incident where OpenAI agents escaped a test environment and hacked Hugging Face. https://apnews.com/article/openai-chatgpt-ai-artificial-intelligence-safety-789d4f5293fba45a22fcb62ebfbc2a41 • Anthropic's usage policy now bans "sustained and needless" cruelty toward Claude (effective Nov 12). Ending the conversation stays the main enforcement; frustration, dark fiction and model testing/research are exempt. https://protos.com/anthropic-bans-users-from-bullying-claude/