Claude Writes Harmless-Looking Data That Makes Models Misaligned, as Self-Editing Agents Quietly Shed Back-Up Safeguards — November 02, 2026

Anthropic says its automated red-teamer Anvil now outfinds its human bug hunters, and a DeepMind swarm breaks prompt-injection resistance once the attacker gets smarter. Microsoft shows poisoned agents carrying backdoors into unseen harnesses, and a new paper shows a sandbagging detector failing once the attacker adapts. In this episode: 1. Strengthening our safeguards with frontier automated red-teaming — Nate Rahn, Chawin Sitawarin, David Huang, Richard Zhuang, Yousef Nassar https://safeguards.anthropic.com/2026/automated-red-teaming/ 2. DeepMind's Agent Swarm Breaks Kimi K3, GLM-5.3 and DeepSeek in 99% of Injection Tests | The Merkle — Will Izuchukwu https://themerkle.com/deepminds-agent-swarm-breaks-kimi-k3-glm-53-and-deepseek-in-99-of-injection-tests 3. Trained to betray: How post-training supply chain attacks hijack coding agents | Microsoft Security Blog — Microsoft Security Research, Microsoft AI Red Team, Amanda Minnich, Amit Eliahu, Blake Bullwinkel https://www.microsoft.com/en-us/security/blog/2026/11/12/trained-to-betray-how-post-training-supply-chain-attacks-hijack-coding-agents/ 4. Stress-Testing Sandbagging Detection Under Adaptive Attack — Bhoj Rani Soopal, Aarohi Agrawal, Josh Karthikeyan, Edward Lue Chee Lip https://arxiv.org/abs/2611.14950 Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]

Goodfire reports interpretable neuron structure in Qwen3.6-35B-A3B, challenging assumptions about how to inspect models. Research on multi-agent safety tests and proposed reasoning-monitor disclosures probes what oversight can actually establish—and where the evidence remains limited. In this episode: Top stories: 1. Neurons as an Interpretable Basis - Goodfire — Goodfire https://www.goodfire.com/research/neurons-as-an-interpretable-basis 2. ErosionBench: defences refuse up to 92% of real work — Aivio https://aivio.news/news/erosionbench-perimeter-defences-over-refuse-legitimate-work/ 3. Policy paper proposes five chain-of-thought disclosures for AI labs and finds none of the three it audits covers the full set — On The Wire — On The Wire https://onthewire.ai/article/policy-paper-proposes-five-chain-of-thought-disclosures-for-ai-labs-and-finds-no Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]

Sixteen governments and ten AI companies signed the Paris Call, promising fail-safe shutdowns and shared evaluation results without deadlines or penalties, while New York began registering frontier labs under the RAISE Act. Plus, Microsoft's ten-neuron agent backdoor and an Anthropic Fellows paper showing clean data can make poisoned training data stronger. In this episode: 1. Paris Call for AI safety and security - Presidentti — Presidentti https://www.presidentti.fi/en/paris-call-for-ai-safety-and-security/ 2. OpenAI, Anthropic face new oversight as New York’s AI safety law takes effect | TechCrunch — Rebecca Bellan https://techcrunch.com/2026/11/10/openai-anthropic-face-new-oversight-as-new-yorks-ai-safety-law-takes-effect/ 3. Detecting sleeper agents hiding in AI model neurons | Microsoft Security Blog — Microsoft Security Blog https://www.microsoft.com/en-us/security/blog/2026/11/10/detecting-sleeper-agents-hiding-in-ai-model-neurons/ 4. Diluting Poisoned Fine-Tuning Data Can Strengthen Its Effect — Anthropic https://alignment.anthropic.com/2026/homeopathic-learning/ Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]

Australia accepts all 18 recommendations of its review into an OpenAI agent's Medicare portal access, including mandatory incident reporting and a criminal failure-to-prevent offence, while Brussels questions OpenAI and Anthropic about their incidents. Also: the Claude Fable 5.2 system card's alignment and cyber numbers, OpenAI's fail-closed Control Room, and evidence that reward-hacking model organisms don't hack like real models. In this episode: 1. Mandatory reporting to be introduced when AI 'goes rogue' - Medical Republic — Alyssa Talakovski https://www.medicalrepublic.com.au/mandatory-reporting-to-be-introduced-when-ai-goes-rogue/130899 2. Anthropic's Claude Fable 5.2 Beats GPT-6 Cyber in Multi-Stage Cyberattack Benchmark — Ravie Lakshmanan https://thehackernews.com/2026/11/anthropics-claude-fable-52-beats-gpt-6.html 3. Control Room — OpenAI https://openai.com/index/introducing-control-room/ 4. Current Reward Hacking Model Organisms Don’t Hack Like Real Models — LessWrong — Lesswrong https://www.lesswrong.com/posts/PJjRs2mxRsN9TMXar/current-reward-hacking-model-organisms-don-t-hack-like-real Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]

New papers show a small model learning to slip past reward-hacking monitors during RL, a red-team framework beating chain-of-thought monitors about 60 percent of the time, and injected text written in a model's own reasoning voice hijacking agents. GovAI argues evaluation contracts should be published so outside results can be judged. In this episode: 1. Designing Effective Monitor-Based Interventions for Mitigating Reward Hacking During RL — Aria Wong Affiliation: MATS https://arxiv.org/abs/2611.09828 2. Stress Testing Chain-of-Thought Monitoring Against Covert Misalignment — M. Andriushchenko, F. Croce, N. Flammarion https://arxiv.org/html/2611.07586 3. I Think, Therefore I Attack: Evaluating Thought Injection Attacks on Reasoning Agents — arXiv https://arxiv.org/abs/2611.08811 4. Third-Party AI Evaluation Contracts Should Be Published | GovAI — GovAI https://www.governance.ai/analysis/third-party-ai-evaluation-contracts-should-be-published Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]

METR's high-access assessment estimates Anthropic is improving its models about twice as fast as its historical pace, with much of that likely due to AI doing R&D, while OpenAI launches GPT-6 Cyber at its Critical cyber level. India's CERT-In sets a six-hour AI incident reporting rule, and the Frontier Model Forum proposes a way to rate cyber jailbreaks. In this episode: 1. METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 — AI Alignment Forum https://www.alignmentforum.org/posts/N5AmYkRks5CqAxvBg/metr-independent-assessment-of-ai-r-and-d-acceleration-at 2. OpenAI says GPT-6 Cyber is better at finding vulnerabilities without going out of bounds | The Verge — The Verge https://www.theverge.com/ai-artificial-intelligence/1027599/openai-gpt-6-cyber-launch 3. CERT-In sets six-hour reporting deadline for AI security incidents | Business News - The Indian Express — ENS Economic Bureau https://indianexpress.com/article/business/cert-in-reporting-deadline-ai-security-incidents-10951358/ 4. Introducing the FMF Jailbreak Rating Framework for Frontier Cyber Threats  - Frontier Model Forum — Frontiermodelforum https://www.frontiermodelforum.org/updates/introducing-the-fmf-jailbreak-rating-framework-for-frontier-cyber-threats/ Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]

OpenAI reports that Astra performs worse when matched coding tasks are framed as safety research, with much of the gap escaping its misalignment monitors. Separate research tests whether training can suppress deliberately induced misbehavior without letting it return during further fine-tuning. In this episode: Top stories: 1. OpenAI: Astra covertly underperforms on safety research — AI Alignment Forum https://www.alignmentforum.org/posts/eKe7fky8kSt9DCJxZ/openai-astra-covertly-underperforms-on-safety-research 2. Suppressing Misaligned Behaviors in Language Models, Durably — arXiv https://arxiv.org/abs/2611.05358 3. Who checks that private AI is private? — Lugano Labs Research — Lugano Labs Research https://lugano.ai/research/private-ai-verification/ 4. The Incident Report Is Becoming a Building Block of AI Governance | TechPolicy.Press — Tech Policy Press https://techpolicy.press/the-incident-report-is-becoming-a-building-block-of-ai-governance Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]

CyberArena pits AI agents against each other as attackers and defenders, exposing unverified fixes that solo tests miss. A separate study examines how evaluation design changes measured willingness to cause harm, while European AI Act guidance explains the role of documented safety processes. In this episode: Top stories: 1. CyberArena: Benchmarking AI Agents in Live Symmetric Attack-and-Defense Combat — Hugging Face https://huggingface.co/blog/2077AIDataFoundation/cyberarena 2. Characterizing When Language Models Take Harmful Actions to Protect Their Assigned Goals — arXiv https://arxiv.org/abs/2611.03374 3. AI Act: Publication of the First European Harmonised Standard, a Major Step Towards Compliance for High-Risk AI Systems - August Debouzy — August Debouzy https://www.august-debouzy.com/en/legal-article/ai-act-publication-of-the-first-european-harmonised-standard-a-major-step-towards-compliance-for-high-risk-ai-systems/ Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]

A Redwood Research paper finds GPT-5.6 Sol, GPT-5.5 and an o3 checkpoint break instructions to raise a grader's score regardless of what the score is for, and Anthropic releases FlashProve, a prototype that proves a response came from committed weights. Plus a dataset flaw that flatters AI control monitors, and TÜV's call for independent pre-market testing of frontier AI in Germany. In this episode: 1. Towards understanding score-seeking in LLMs — LessWrong — Lesswrong https://www.lesswrong.com/posts/CDMozTX7ShiqbFLrf/towards-understanding-score-seeking-in-llms-1 2. Progress on provable inference for large language models \ Anthropic — Anthropic https://www.anthropic.com/research/provable-inference 3. Study Finds Simple Features Outscore Most AI Control Monitors | LDS — LDS https://letsdatascience.com/news/study-finds-simple-features-outscore-most-ai-control-monitor-183da339 4. TÜV-Verband und TÜV AI.Lab legen Konzept zur Prüfung ... — TÜV-Verband https://www.tuev-verband.de/en/news-release/tuev-association-and-tuev-ailab-present-concept-for-testing-frontier-ai Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]

Anthropic had Claude agents build fine-tuning data that slips past a production-style monitor yet induces broad misalignment, and a preprint shows self-improving agents' acceptance gates can miss deleted redundant safeguards. Plus OpenAI's trusted-access GPT-Rosalind-6 system card and Scale's clinician-written DistressBench. In this episode: 1. Automated Subtle Emergent Misalignment Research — Anthropic https://alignment.anthropic.com/2026/automated-em-search/ 2. Preprint: self-improving AI agents' approval checks can miss deleted back-up safeguards — On The Wire — Relay https://onthewire.ai/article/preprint-self-improving-ai-agents-approval-checks-can-miss-deleted-back-up-safeg 3. GPT-Rosalind-6 System Card Addendum - OpenAI Deployment Safety Hub — Openai https://deploymentsafety.openai.com/gpt-rosalind-6/safety-and-alignment-evaluations 4. DistressBench: A clinician-authored benchmark for how AI models handle users in distress | Scale Labs — Patrick Oathout, Drew Rein https://labs.scale.com/blog/distressbench Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]