Booz Allen Hamilton says its OT lab ran eight attack scenarios with two unnamed frontier models, and the models met the objectives in all eight. One agent moved a robotic arm within minutes; another reached an industrial control network in just over 16 minutes. Kyle Miller, a VP at the firm, warns defenders may have minutes to respond.
DeepSeek has agreed to lead the seed round for Winverse AI, a robot AI start-up founded by former Tencent scientist Hu Han, Bloomberg reports, citing people familiar. The round would value Winverse at about $400mn, with Meituan's Long-Z also committed. The deal is not final.
Anthropic's updated usage policy, effective November 12, prohibits sustained and needless abusive or cruel behavior toward its Claude models, while tightening restrictions on election interference, weapons development and surveillance, the company says. It dropped its blanket ban on personalized political targeting. Anthropic says Claude can still end abusive conversations.
SemiAnalysis finds that of 857 model releases from nine Chinese labs between September 2021 and September 2026, only 31, about 3.6%, carried safety reports. Alibaba shipped 238 models with seven reports. The firm says enterprises adopting cheaper Chinese open-source models for cost and control take on the safety burden themselves.
Anthropic CEO Dario Amodei pledges to embed independent evaluators inside his company, and OpenAI CEO Sam Altman endorses the idea. Nonprofit evaluator METR reports about $71 million in recent funding commitments. OpenAI fired three employees, which two say stemmed from contact with third-party evaluators; OpenAI disputes that and says it is finalizing assessor contracts.
Microsoft CEO Satya Nadella argues in a Saturday essay that AI models pose insider risks and should be governed with zero trust principles, including tamper-proof audit trails and independent validation. SemiAnalysis separately reports that only 31 of 857 model releases from nine Chinese labs, about 3.6%, included safety reports.
Ryan (@wheresryan22) releases Anatomy, an agent skill that turns technical ideas into interactive isometric SVG figures of invented machines, such as a token bucket rate limiter drawn as a pin-tumbler lock. It runs in Claude Code, OpenCode and Codex and installs with npx skills add wheresryan22/anatomy. Node 20.10 or later is required, and the 3D mode is beta.
Humanbound ships v2.13.0 of its open-source adversarial testing engine, SDK and CLI for AI agents, licensed Apache-2.0. The company says hb guardrails turns failed test findings into deployable firewall rules, and hb posture scores an agent's security from 0 to 100. Runs locally or against the Humanbound Platform.
AirLLM's v4.0.0 release adds training support for Qwen3.8-Flash-Next, a 125B model, under 6GB of VRAM, streaming frozen weights one decoder layer at a time while adapters stay on the GPU. The project, maintained by Gavin Li, says it also runs Kimi K3 (2.8T) in 3.72GB. The Apache-2.0 library installs via pip.
Sentry engineers Ryan and Serge show how the company traces its own AI agents in production, using Agent Tracing to collect traces, investigate unexpected behavior, and surface failures too numerous to review individually. They describe using signals such as user frustration to find problems at scale. The session is part two of a workshop series on debugging agents.