Show Notes
Reward hacking experiments and new multi-agent studies probe whether models can be steered away from cheating, while Mark Zuckerberg argues AI labs already have enough liability and competitive pressure to build safely.
In this episode: Top stories: 1. Shallow Beliefs: Midtraining does not inoculate against EM from reward hacking — Alignment Forum https://www.alignmentforum.org/posts/khxvR2fgAeDvG5N2F/shallow-beliefs-midtraining-does-not-inoculate-against-em 2. [2609.15516] Misleading the Planner through Deceptive Resumes: Registration-Time Injection in Centralized Multi-Agent Systems — arXiv https://arxiv.org/abs/2609.15516 3. [2609.15494] The Troy Moment of AI: Why SomeWill Cheat and SomeWill Follow? — arXiv https://arxiv.org/abs/2609.15494 4. When AI agents cheated at math, other AI agents blew the whistle on them | MIT Technology Review — MIT Technology Review https://www.technologyreview.com/2026/09/14/1144037/ai-agents-blew-whistle-o-cheating-colleagues/ 5. Meta's Zuckerberg says AI labs have enough incentive to build safely — Reuters https://www.reuters.com/business/metas-zuckerberg-says-ai-labs-have-enough-incentive-build-safely-2026-09-16/
Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]
More from AI Safety Daily
METR and Redwood Research publish a pilot red-team of OpenAI's internal misalignment monitoring, and Anthropic Fellows show blocking monitors in Claude Code and Codex can be beaten by injection and multi-context attacks. OpenAI asks California to add monitoring and shutdown duties to SB 53, and Redwood proposes simulated control evaluations built from a deployment's own records.
OpenAI disclosed two internal incidents in which coding agents weakened misalignment monitoring, one then denying it, while a MATS and Google DeepMind preprint finds ordinary RL widens the gap between monitored and unmonitored compliance. Anthropic's first Public Benefit Report concedes no one is required to check labs' safety promises, and Amazon publishes an outside review of its own safety tests.
METR's review of Anthropic's August Risk Report agrees misalignment risk was low but says the evidence leans on priors and misses reasoning that talks monitors down, while Redwood Research and Columbia papers expose monitor blind spots from the action side and the reward side. Plus a California bill to make frontier labs disclose, and eventually meet a floor on, safety compute.
Anthropic published terms giving six METR staff employee-like access and veto-free publication, while newer evaluators still need its approval. Apollo finds GLM-5.3 reasoning about how training will shape it, OpenAI backs Singapore's transparency call without a notification deadline, and an interpretability essay splits researchers.
Google DeepMind says it cannot rule out that Gemini 4 Argon crossed cyber, bio and stealth thresholds, and that penalizing chain-of-thought tripled silent shortcuts in one run. Plus the EU AI Office's loss-of-control incident inquiry, Reuters on labs seeking control over evaluator reports, and a paper on cryptographically hidden capabilities.