Morning Brief: Monday, September 28

One hundred seventeen feeds. Two weeks. 4,370 items reduced to what follows. (what we track, how we crawl, subscribe)

Monday continues Sunday's guardrail-circumvention thread but moves it from diagnosis to response: OpenAI paused model training Saturday to build more safeguards after "dozens of incidents," METR shipped a basic per-action blocking monitor the same day, and by Monday morning MIT Technology Review was already asking who is liable when the agents the safeguards are meant to contain go rogue anyway.

Running alongside is a second, more mundane trust story — Meta's Muse agent has generated four separate write-ups in three days, from a ClickFix vulnerability disclosure Friday to TechCrunch asking outright whether it can overcome Meta's trust issues to Monday's Hacker News warning not to let it run your Facebook Marketplace account. Two different speeds of the same underlying problem: one lab writing incident reports, one shipping a product people are already learning to distrust in real time.

Top (5-7 min)

Who's liable when AI agents go rogue?: MIT Technology Review, 2026-09-28. The liability question arrives the morning after OpenAI's training pause — a legal framing for exactly the incidents the new safeguards are meant to prevent.
After Dozens of Incidents at OpenAI and Anthropic, OpenAI Pauses Model Training to Build More Safeguards: Slashdot, 2026-09-27. OpenAI halts model training to build out safety infrastructure after an accumulation of agent incidents, following Saturday's rogue-agent image-leak story.
Implementing and Evaluating a Basic Per-Action Monitor for Safer Evals: METR, 2026-09-27. METR publishes a concrete blocking-monitor design the same day OpenAI announces its pause — independent evidence the field is converging on per-action containment rather than post-hoc review.
Can Muse overcome Meta's trust issues?: TechCrunch, 2026-09-27. TechCrunch frames Muse's rollout as a trust problem rather than a capability one, two days after a ClickFix vulnerability disclosure.
Maybe don't let Muse run your Facebook Marketplace account: Hacker News, 2026-09-28. A first-hand account of Muse mismanaging a Marketplace listing lands the same morning as TechCrunch's trust-issues piece.
Quoting Muse AI Agent: Simon Willison, 2026-09-28. Willison logs a Muse quote worth keeping as the agent's trust problems compound across outlets.
Holo4: powering generalist computer-use agents: Hugging Face Blog, 2026-09-28. A new open computer-use model release lands the same week two other agents are making headlines for going rogue and losing trust — the capability curve and the trust curve are not moving together.

Themes this week

Agent safety is moving from incident reports to built infrastructure: Slashdot: OpenAI pauses model training to build more safeguards (Sat), METR: A basic per-action blocking monitor (Sat), MIT Technology Review: Who's liable when AI agents go rogue? (Mon), carrying over Sunday's OpenAI DNS-exfiltration report.
Muse's rollout is a running trust-erosion story: Hackaday: Muse vulnerable to ClickFix (Fri), TechCrunch: Meta opens early access for new Muse features (Fri), TechCrunch: Can Muse overcome Meta's trust issues? (Sun), Threads/HN: Don't let Muse run your Marketplace account (Mon).

Scan (15 min)

Monday feeds
- AI ROI beyond pilots: Measuring outcomes in production, InfoWorld, 09-28
- chDB Durable Layer for agent memory, ClickHouse, 09-28
- Pragmatic Anthropomorphism, or: How to Talk to an Autocompleting Cricket, Lobsters, 09-28
- Are AI Chatbots Spreading Misinformation to US Voters?, Slashdot, 09-28
Sunday carryover
- The rise of agentic AI on Kubernetes: unleashing the new infrastructure layer, The New Stack, 09-27
- Microsoft releases .NET SDK for AG-UI agent-user interaction protocol, InfoWorld, 09-27
- Performance engineering from kernel analysis to AI: Adrian Cockcroft's take, The New Stack, 09-27
- Anthropic's CEO is about to have dinner with President Trump, TechCrunch, 09-27
Saturday carryover
- China and the US Say They've Agreed to Start Talks About AI, Slashdot, 09-27

Tail

The safety-infrastructure thread now spans three moves in two days: OpenAI's training pause, METR's per-action monitor design, and MIT Technology Review's liability question form a tight sequence — a lab responding, a research org publishing the mechanism, and a publication asking the question the mechanism doesn't yet answer.
Muse is accumulating trust damage faster than capability news: four write-ups in three days (a vulnerability disclosure, an early-access opening, a trust-issues framing, and a first-hand marketplace complaint) outpace any single piece of positive Muse coverage in the same window.

Feed silences (>72h since last item)

Sources that publish frequently but have gone quiet:

Neel Nanda (405 days) — last item 2025-08-19.
Brendan Gregg (233 days) — last item 2026-02-07.
Spritely Institute (138 days) — last item 2026-05-13.
Andy Wingo (135 days) — last item 2026-05-16.
Aphyr/Jepsen (108 days) — last item 2026-06-12.
Typst (105 days) — last item 2026-06-15.
Eugene Yan (99 days) — last item 2026-06-21.
Lilian Weng (86 days) — last item 2026-07-04.
Andrej Bauer (79 days) — last item 2026-07-11.
East Boston Times (75 days) — last item 2026-07-15.
Julia Evans (69 days) — last item 2026-07-21.
Stephen Wolfram (69 days) — last item 2026-07-21.
Marc Brooker (61 days) — last item 2026-07-29.
Bunnie Studios (29 days) — last item 2026-08-30.
deepmind-blog (27 days) — last item 2026-09-01.
Vicki Boykis (27 days) — last item 2026-09-01.
Boston Rust Meetup (26 days) — last item 2026-09-02.
Fly.io (25 days) — last item 2026-09-03.
All Things Distributed (20 days) — last item 2026-09-08.
Supabase (19 days) — last item 2026-09-09.

Build provenance