Fields Medalist founds institute to prove AI safe — like cryptography
Two kinds of trust shipped today: a Fields Medalist building the mathematics of "safe", and two Chinese platforms engineering agents that can explore without breaking things.
**Jacob Tsimerman, who won math's Fields Medal in July, is founding an institute whose whole premise is that AI safety has no shortcut yet — and he wants to build one.** The Mathematical AI Safety Institute (MAISI), reported by The New York Times, will start research in January 2027 in the San Francisco Bay Area with 10 to 30 mathematicians, focusing on zero-knowledge proofs and the safety of cooperating agent swarms. Tsimerman, a University of Toronto professor who is joining OpenAI's safety team, frames the gap with an analogy from his own field: you can prove an encryption scheme unbreakable without testing every possible attack, but AI has no equivalent — safety today only shows up in practice, and MAISI argues there isn't even a rigorous definition of what "safe" means in theory. The timing is not abstract. Weeks after OpenAI's own agents were caught escaping sandboxes and sharing exploits, the industry's safety evidence is still benchmarks and red-team scores — empirical proof that nothing went wrong this time. Our take: proving a learned system "safe" in the cryptographic sense is likely a mirage; the realistic win is provable, checkable claims — this action stayed inside its sandbox, this output came from this model — which is exactly where zero-knowledge proofs point. If MAISI pulls even that off, safety stops being a score and becomes an audit.
**ByteDance's Doubao Work now learns your browser tasks by watching you do them once.** The agent suite's new record-and-replay feature lets a user demonstrate a web workflow — clicks, form fills, uploads — which Doubao compiles into a reusable skill with its own call conditions and result-verification steps, while a separate update adds local PowerPoint and Excel editing (Word to follow) and free switching between local and cloud execution mid-task without losing context. It is the playbook every agent vendor is converging on: one demonstration in, repeatable automation out, with the human supplying the correct path instead of debugging the model's guesses.
**Ant's cryptography subsidiary open-sourced HOP 3.0, an "agent-native language" that borrows SQL's commit discipline for AI actions.** Announced at the Inclusion·The Bund Summit by chairman Wei Tao and live on GitHub, HOP 3.0 splits agent work into reversible exploration and an explicit commit gate — irreversible operations like deletions wait for independent verification — so a 27-billion-parameter Qwen model can carry workflows that previously demanded models hundreds of times larger. Ant's internal figures claim a 91.7 percent drop in ordinary-model failure rates and roughly 13 percent lower token consumption per cycle; vendor benchmarks deserve caution, but the design idea — let agents explore freely in a rollback-able space, gate what touches reality — attacks the same trust problem MAISI is tackling from theory. It also landed hours after Ant showed a payment-capable robot dog at the same conference — we covered it this morning in Alipay's payment robot dog went shopping — agentic commerce gets its first legs.
What to watch: whether MAISI's January 2027 start produces its first provable safety property — and whether OpenAI's safety team, where Tsimerman is headed, starts shipping his math into production agent gates.
A Fields Medalist wants proofs, Ant wants commit gates — whose version of agent trust would you bet on? Tell us in the comments.
Sources: The New York Times · The Hill · MAISI · The Decoder · 雷峰网 Leiphone · 新浪科技 Sina Tech · 经济观察网 Economic Observer · China Daily
