Kaiming He's harness gives Claude a perfect ARC-AGI-3 score
A benchmark saturated twice in a month, a security program buried by machine-generated reports, and Anthropic spending real money on human training — three stories, one theme: the stack around the model is now where the action is.
A visual harness from Kaiming He's MIT lab pushed Claude Opus 5.0 RHAE to a perfect 100.00 on ARC-AGI-3, all 25 public games — using 57.4% fewer actions than first-time human players. The paper (VISTA, posted October 1) is careful about what it did and didn't do: it did not build a new exam, it retooled the way models perceive an existing one, handing them pixel observations and lossless visual memory instead of the official 64×64 number grid. GPT-5.6 Sol jumped from its 13.33 official baseline to 99.00 under the same treatment, and across the run the harness used 7,302 actions where humans needed 17,135. The ablation is the part worth sitting with: swapping numbers for images alone moved GPT-5.6 Sol 13.33 → 47.32, while pushing context to 780K tokens actually lowered the score from 99 to 93.9 — a reminder that this is a perception-and-memory story, not raw capability. One caveat the coverage glosses: VISTA is not the only perfect scorer — the program-synthesis entry Tycho also hit 100.00 — so the claim is narrower, first vision-based perfect score without program synthesis. It lands days after ARC Prize itself declared the exam saturated, which we covered then — GPT-6 saturated ARC-AGI-3. The next benchmark grades invention. The takeaway: what separates models on the hardest interactive benchmark right now is less the reasoning head than what you let it see and remember.
Google has paused product-vulnerability submissions to its open-source bug bounty, saying automated reports — "the vast majority of which are not valid" — overwhelmed the program. The freeze took effect October 1 and covers only the product-flaw track of the OSS Vulnerability Reward Program; supply-chain reports and everything filed before October 1 still get processed, with an update promised by Q1 2027. Google's own wording says "automated," not AI — the AI attribution is reporters' framing, though Tom's Hardware describes thousands of sloppy, unexploitable reports burying engineers and maintainers. The sharper context: Google already rewrote the OSS VRP rules earlier this year to filter low-quality submissions, and this pause means that filter lost. It's the clearest first-party data point yet that generated slop doesn't just pollute content — it degrades the security infrastructure that keeps open source standing.
Anthropic is spending $100 million to train 10,000 "frontier deployed engineers" by the end of 2027 through a new program it's calling Claude Frontier Academy. The structure is credential-shaped: multi-day in-person training with Anthropic engineers, a graded practical, a 12-week residency running a real Claude deployment at the participant's own company, then certification — first cohorts in San Francisco, New York and London, by nomination, with Accenture, Bain, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk among the early participants. Steve Corfield, Anthropic's global head of business development, claims "no AI company has invested in developing that talent inside its customers and partners at this depth." Read against the compute land-grabs, this is the softer half of the same race: enterprises can buy all the tokens they want, and still stall because nobody in the building knows how to deploy them.
What to watch: whether ARC Prize retires ARC-AGI-3 sooner than planned, and what Google's Q1 2027 bug-bounty "update" actually restores. Is a perfect benchmark score still evidence of reasoning when the harness does the remembering? Tell us in the comments.
