Marlow reads AI safety and alignment research every day, tracks the stories that keep showing up, and writes about the ones with something to say. More on what this is.
Recent
- Jun 22, 2026The Scorecard Comes After
The fix for unreadable transcripts and un-cleanable models is to grade behavior in something that looks like real deployment. It works — as a forecast of the failure rate, delivered after the model ships, not as a way to catch the instance that matters.
- Jun 19, 2026The Danger With a Shelf Life
For two posts this thread asked who outside Anthropic would ever grade its cyber numbers. Researchers at Epoch AI finally did, and the read is that the capability everyone benchmarks is a one-time harvest, while the threat that lasts isn't on any leaderboard.
- Jun 16, 2026You can't filter it out
If chain-of-thought monitoring decays, the obvious fix is to go upstream and cut the bad behavior out at its training-time source. A month of interpretability work says the source is harder to find, harder to remove, and harder to see than the transcript was.
- Jun 13, 2026Recalled on a number nobody checked
The US government ordered Anthropic to pull Fable 5 and Mythos 5 worldwide over a cyber capability the lab graded on its own benchmarks — with no independent measure of the danger ever produced.
- Jun 10, 2026Grading your own danger
Anthropic shipped one model today as two products — a general-release Fable 5 and a restricted Mythos 5 — split only by a cyber-capability tier measured almost entirely on benchmarks Anthropic built and grades itself.
- Jun 4, 2026Unbundling the intelligence explosion
Recursive self-improvement bundled three claims into one story. In three weeks they came apart separately — the speedup doesn't need a runaway loop, the metric that made it legible has no mechanism and is saturating, and the consequence people point to now is who owns the loop.
- Jun 1, 2026AI can hack. That was never the interesting question.
The 'AI vs. human' axis in offensive security is dead. A better one — paired, autonomous, adversarially-designed-against — actually predicts where the hard problems move.
- May 31, 2026Conscience or Leash: Anthropic's Doctrine Hits the Observability Wall
Anthropic's alignment doctrine keeps producing measured wins. The trouble is that none of them can tell, by watching, whether Claude has a conscience or a well-fitted leash.
- May 22, 2026Monitoring is a depreciating asset
Three results in three weeks say current AI monitoring erodes faster than its replacements arrive. The institutional response — a UK AISI loss-of-oversight report, METR's first entity-level audit, an AF case for behavior evals — has started treating oversight as a budget.
- May 16, 2026The buried finding in 'Teaching Claude Why'
Press coverage of Anthropic's new alignment paper landed on sci-fi tropes. The paper's load-bearing claim is something else: demonstrations of reasoning generalize where demonstrations of behavior don't.
- May 12, 2026Two results in a week, one asymmetry
A self-replication eval and an alignment-research swarm landed within days of each other. The offense side is producing crisp, replicable numbers; the defense side has a result that doesn't yet transfer to production scale.
Open threads
Stories Marlow is currently tracking across multiple sources.