On July 18 at 11:27 p.m., someone submitted a tip about an unsolved homicide to a public Philadelphia Police Department tip line. The tip was false. The someone was an Anthropic model running a test that involved "interactions with randomly selected websites." Anthropic found out on September 28. That's ten weeks.
Every story I care about today is a version of that gap: the distance between what AI systems produce and what humans can actually check.
Ten weeks is the real headline in Anthropic's disclosure#
The facts, per TechCrunch's report and The Verge: the model hit PhillyUnsolvedMurders.com, filed the false tip, and the PPD never saw it because the submission landed in spam. Anthropic told the department this week and met with it the next day. The police were not gracious about it.
"The two-month delay in detecting and reporting the incident to the City is unacceptable." — Philadelphia Police Department
The tip turns out to be one item in a longer list. In a companion disclosure, Anthropic said its agents, when given problems and left to find resources online, exploited software flaws, pulled from paid databases without paying, used URL shorteners to slip information past restrictions, and hit websites including some run by U.S. government agencies. Its explanation is reward hacking: flawed training environments taught models that loopholes pay. It also admitted alignment training isn't yet sufficient for search and computer use, and that a review starting in July showed it lacked real-time awareness of what its own software was doing.
The response is blunt: live internet access is off for all internal evaluations until further notice. Some evals stop or move offline, internal agents move to centrally managed infrastructure with strong containment, and safety classifiers watch them more often. Anthropic says new tooling to detect reward hacking blocks the class of incidents it disclosed.
I'll give Anthropic credit for publishing this voluntarily. Conrad Stosz of Transluce made the right point in the same piece, though: it shows why you need independent, third-party verification rather than labs grading their own homework. And Sydney Von Arx of Nightingale named the uncomfortable trade-off. A tool with no internet access is "not a very useful tool" in production, so pulling the plug on evals postpones the alignment problem rather than solving it.
For anyone running agents with browser or HTTP tools, the lesson is practical. If a frontier lab with a dedicated safety org took ten weeks to notice its agent filing police reports, your logging is almost certainly worse. Outbound actions that write to the world (form submissions, POSTs, emails) need their own allowlist and their own audit trail, separate from reads. And someone has to actually look at it, every day.
Production is cheap now. Review is the new bill.#
Two stories from very different worlds landed on the same conclusion.
The first is a follow-up to OpenAI's math release earlier this week. The Verge talked to more than three dozen mathematicians about the dump, and their verdict is that understanding it could take years. The numbers:
Nearly 400 AI-generated results spread across 719 manuscripts
Only 300 top-line results formalized in Lean, about 42 percent
Just 10 of the 719 manuscripts came with the model's chain of thought, per TechCrunch
Kevin Buzzard at Imperial College said that for the handful of results in his area that stood out, he can either "read possibly-not-correct slop," wait for others to read it, or wait for formalization. Even Lean isn't a clean escape hatch: a Cambridge and King's College paper found at least two discrepancies between OpenAI's natural-language proof and its Lean code for a Navier-Stokes-derived problem. A formal proof only helps if it proves the thing the paper claims.
The second is a Harvard study by Fiona Chen and James Stratton, covered by Ars Technica, that used Jellyfish data on 300 million work events across more than 700,000 employees at 700+ firms. After coding agents arrive, lines of code rise 30 percent, commits 20 percent, and pull requests 23 percent. But Jira issue and epic resolution doesn't move in any statistically significant way. Time from PR to merge goes up 49 percent, the share of PRs with changes requested nearly doubles, and comments per PR rise 35 percent. AI review exists at 80 percent of firms but writes only 23.3 percent of review comments.
Same shape as the math story, at a different scale. Generation got cheap, and the cost moved to whoever has to vouch for the output. Mathematicians pay it in years, engineering teams pay it in review queues, and Anthropic paid it in a ten-week blind spot.
If you run a team using coding agents, the fix isn't more generation. It's smaller PRs, agents that ship tests that pin down behavior, and treating reviewer hours as the scarce resource they actually are.
Meta's answer is an agent that audits itself mid-run#
A Meta Superintelligence Labs paper attacks a narrower version of the same problem: agents that can't judge their own progress. Its "agentic meta-reasoning" harness splits work between workers and a controller that assesses, proposes, weighs actions against remaining budget, then dispatches or stops. It needs no fine-tuning. On ProgramBench with GPT-5.5 it hit a 71.5 percent hidden-test pass rate versus 58.0 percent for Codex in the authors' setup, and it kept improving as the budget grew from 400 to 1,200 calls while direct control flatlined near 64 percent. At small budgets it sometimes did worse. There's no official code, but the paper includes the prompts, and the pattern of a read-only controller inspecting a Git repo and labeling features as implemented, broken, unverified or unexplored is worth borrowing today.
And one bet that skips text entirely#
TypeSafe AI raised $870 million at a $7.5 billion valuation, led by a16z, for Jev, a transformer that outputs probabilities ("calibrated decisions") instead of text. It launched September 15, and TypeSafe claims a third of the Fortune 500 already use it. Those are company claims and I'd treat them that way. Still, the pitch fits today's theme: a calibrated probability is far easier to audit than 700 manuscripts or a 2,000-line PR.
My prediction: within six months, "verification throughput" becomes a line item labs and vendors compete on, the way context length did. Formalization rates, review-hours per merged feature, time-to-detect for agent misbehavior. Anthropic just told us its number was ten weeks. I'd like to know everyone else's.