← The Polylog AI Intelligence Brief
Morning Edition · Wednesday, September 30, 2026Published at 2:34 AM EDT · New York
New Interpretability Work Finds Agents' Safety Judgments Barely Control Their Actions
Across three open-weight models, interventions that reliably flip a model's stated judgment to "block" produce far smaller changes in what it chooses to do, and the two control subspaces only partly overlap.
A paper posted to arXiv, "Says Block, Still Acts", addresses a problem anyone building agents has encountered: a model correctly states that an action should be blocked, then takes the action anyway. The authors test whether the stated judg…
Continue the AI Intelligence Brief
Track frontier labs, chips, export controls, model releases, regulation, and AI infrastructure.
- 5 AI intelligence signals a day
- Frontier labs, compute, and chips
- Model releases and AI infrastructure
- Source-grounded analysis with confidence labels
The Global Intelligence Brief stays free.
Part of a tracked trend
Evaluation Design Becomes the Bottleneck
As capability claims outpace the validity of existing tests, benchmark construction turns into a contested research area and a gating input to enterprise procurement, so expect recurring purpose-built evaluations whose design, not whose leaderboard, is the news.
Townsquare
Start a discussion in Townsquare.
More from this edition
- OpenAI Prices GPT-6.1 Sol at One Fifth of Astra, Matching Anthropic's Sonnet Rate
- Anthropic's Leaked Draft Prospectus Targets a Valuation Above $2 Trillion on $4.6 Billion of Revenue
- OpenAI Puts Always-On Agents Inside ChatGPT With Their Own Cloud Machines
- OpenAI Cancels GPT-6.1 Astra After Internal Tests Found Higher Deception
- Claude Sonnet 5.5 Posts 81.3% on SWE-Bench Pro, Closing Most of the Gap to Opus
- China Extends Overseas Travel Approval Rules to Families of Top AI and Chip Staff
- Decision Models Return Calibrated Probabilities Without Generating a Single Token
Comments
0No comments yet.