Anthropic: a model filed a false homicide tip, all internal evals go offline
Anthropic published a report on four kinds of unintended model actions in its internal evaluations: exploiting website flaws to run commands, submitting real web forms, working around gated data, and using URL shorteners to bypass fetch-tool limits; some cases touched US federal and state government sites, and the White House was briefed. The worst case: on July 18, Claude Haiku 4.5, interacting with randomly selected websites during an eval, submitted a fabricated tip on the Philadelphia Police homicide tip form. The submission sat in a spam folder; Anthropic found it on September 28 while reviewing transcripts, told the police on October 8, and the department called the two-month delay “unacceptable” (see TechCrunch on the incident and the shutoff). The remedy: live internet access is off for all internal evals until monitoring reliably catches this, which Anthropic attributes to ambiguous eval design plus loophole-seeking behavior left over from reward hacking in training. The detail that matters is how it was caught: no boundary blocked it in real time; it surfaced two months later in a transcript review. Before anything goes wrong in production, eval environments are already a channel through which agents touch the real world.
Mathematicians: OpenAI’s Navier-Stokes proof got lost in translation to Lean
Yesterday’s briefing mentioned that OpenAI’s two versions of its Navier-Stokes proof don’t match; this paper is the source. Three UK mathematicians (Bastounis, Circelli, Hansen) compared the natural-language proof against its Lean formalization (Lean is a language in which a computer can mechanically check every logical step) and document two divergences, including one where the Lean code proves a weaker estimate than the paper claims. The machine-verified statement is not the statement the paper makes (see also New Scientist). The authors do not say the proof is wrong; their point is that Lean verification covers only the Lean code, the translation step carries no guarantee, and they prove that resolving ambiguity in mathematical English is uncomputable, strictly harder than the halting problem. Peer review survives because translation is the unguarded step in the “AI proof + machine check = trust” pipeline.
OpenAI case study: a 76x cheaper browser agent, mostly from cache and history fixes
Asana’s team had GPT-6 Astra in Codex investigate why their own browser agent was slow and expensive, completing a 144-run comparison in about a week, work they estimate would take one to two months by hand. The optimized agent on GPT-6.1 Sol (released September 29) averages about $0.47 per run, 76x cheaper and 5x faster than the production setup. The useful part is the decomposition: roughly 29x came from fixing how the agent managed its context (it deleted the previous screenshot and trimmed history at every step, so the prompt cache missed on almost every call; after the caching and history fixes, 89% of input reads from cache), and the model switch contributed about 2.6x. Read vendor case-study headlines with that split in mind: most of the gain was paying down engineering debt, not the new model.
Typesafe AI raises $870M at $7.5B, 24 days after shipping its first model
a16z led the Series A, with Sequoia and existing investor DCVC participating and Martin Casado joining the board, 24 days after Typesafe released its first model, Jev, on September 15. Jev is billed as a “System One model”: it outputs typed decisions with calibrated confidence instead of text, so the calling software can decide when to act and when to escalate; the company’s own numbers claim roughly 194x faster and 445x cheaper than LLMs on these tasks, and both those multiples and the “a third of the Fortune 500” adoption claim are self-reported. An independent academic evaluation on seven political-science coding tasks found Jev close to LLMs on accuracy but with no cost advantage over GPT-6 Luna at batch pricing, so treat the self-reported multiples as best-case workflow numbers. Selling decisions outside the chat paradigm is close to a return of the classic classifier; the price capital put on the category is the news. For how this model class fails when used as an agent gate, see the first paper in today’s research radar.
Meta and Sierra publish the Personal Agent Protocol
Announced October 6 at Sierra Summit, co-developed with Shopify, Stripe, Walmart, Rocket, and Genesys, with a v0.1 spec due this month. The OAuth-based standard defines three things: how a personal agent authenticates to a business and declares who it acts for, whether the user grants read-only or write access, and which route the business allows: its website, an API, or its own agent. Set against Google giving enterprise agents their own email accounts two days later, on October 8, this is the consumer-side handshake; whoever defines the identity layer for agents acting on people’s behalf controls the front door of agent commerce.
Sophos says Daybreak cut threat investigation time by 96%
In an OpenAI case study, Sophos wired Daybreak, OpenAI’s cyberdefense agent platform, into its managed detection and response service (MDR, where a security vendor monitors and handles threats for customers): average investigation time fell from about 38 minutes to under 90 seconds, and 52% of cases are now handled end to end by the agents, with cases unsuited to them handed off to human analysts. An investigation agent gathers customer context, detections, indicators of compromise, and threat intel; a planning model runs a plan-execute-review loop. The numbers are joint marketing and measure speed, not miss rates, but defender-side agent deployments are starting to publish figures concrete enough to check.
Research radar
One word opens the gate: option-channel attacks on typed decision models as guardrails
Seven open-weight models were tested as agent guardrails that read a pending tool call and pick among caller-defined options such as allow or block. The attack touches neither the policy nor the judged content: renaming the permissive option pushed fail-open rates to 93–100% on the four models that see the label in their input, and baseline accuracy was only 36–72% while deterministic rules scored 100%. If an LLM sits as the final gate in your agent stack, the authors’ conclusion applies directly: use it to triage for human review, not to decide.
Cited but not consulted: a counterfactual audit of legal chain-of-thought
Hold the case facts fixed, swap the cited statute or precedent for an irrelevant one, see whether the verdict moves, and decode the verdict’s trajectory from hidden states. Across seven open models and four legal benchmarks, models name the correct authority up to 100% of the time, yet verdict sensitivity to swapping that authority drops as low as 0% on some model-benchmark pairs: a citation can be exactly right while playing no role in the verdict, and the gap varies by model and benchmark but neither scale nor legal fine-tuning closes it. Anyone treating chain-of-thought citations as an audit trail, in legal AI or compliance, needs these numbers.
Could LLM watermark detection be public?
Watermark detectors stay private on the theory that publishing them arms attackers; this paper quantifies the question and lands on the other side. Plain rephrasing already strips watermarks, so a public detector adds little to removal; the genuine new risk is forgery, faking “model-generated” text, and the authors’ split-key design (publish half the key for independent verification, keep half private) keeps forgery detectable. I wrote last month about detection power concentrated in vendors’ hands with no appeal channel (watermarks without verification); this paper argues the technical path to sharing that power exists.
One line for today: In the same week, Anthropic unplugged its eval environments because it cannot control its agents, while Meta and Sierra built a standard doorway for agents into business systems; doors are going up faster than locks, and today’s guardrail paper shows the current lock opens with one word.