DeepMind reports machine-checked proofs for selected security properties of Signal’s cryptographic implementation, with humans approving the specifications. Anthropic reports three working electronic prototypes designed with Claude and physically built by engineers. These November 9–10 reports show ways AI can contribute to demanding technical work, but neither establishes complete software security, production-ready hardware or general engineering autonomy.
Sources checked and limits of this review
Review since the November 9 briefing. Checked primary watchlists at METR, ARC Prize, Epoch AI, Google DeepMind, OpenAI and Anthropic, plus independent-research searches. Axios and Wired supplied the Signal-verification lead; claims were checked against indexed excerpts of DeepMind’s original announcement, technical paper and official project page. Direct retrieval of those three primary pages failed, so this note is bounded by the original-source text available in search; no proof replay or full-paper review is claimed. Anthropic’s November 9 electronics report was read directly, including its human roles, commercial relationship and validation limits. Epoch’s index lists an inference-capacity report on November 9, while its hardware changelog dates the underlying index addition to October 19; the accessible methodology was checked, but this was not reported as a newly launched metric. No new item was selected from the METR, ARC Prize or OpenAI watchlists or the independent-research leads checked. This is a bounded review, not an exhaustive claim about all developments. Lab-reported findings are distinct from editorial SI interpretation; no human editorial approval, overall wisdom verdict or pivotal designation is claimed.
Evidence note 1 · Lab-reported formal verification; SI relevance is editorial interpretation
DeepMind reports proofs for selected Signal security properties
DeepMind says five researchers working with AI agents over four weeks verified key properties of a compiled libsignal implementation for one-to-one messaging. Humans guided and approved formal specifications; Lean checked agents’ proofs against them. The work connects protocol-level claims to machine code.
Our interpretation. Editorial interpretation: checking demanding software properties could strengthen defenses as AI-generated code and cyber capabilities expand. A checked proof can constrain plausible errors, but only within a correctly chosen specification and its assumptions.
Main limitation. Lab-reported result, not an independent replay. Trust remains in specifications, machine semantics, cryptographic assumptions and contracts for some routines. Group messaging is outside scope; the computational proof is classical. This does not certify every Signal app or all attacks. Primary text was inspected through indexed excerpts because direct retrieval failed.
Next evidence. Independent proof replays and specification audits, broader protocol coverage, and measured human effort and compute costs on additional codebases.
Evidence note 2 · Lab-and-commercial-partner hardware demonstration; SI relevance is editorial interpretation
Anthropic reports three working prototypes from Claude’s electronics designs
Anthropic says Claude Fable 5.2 designed a speakerphone, voice-assistant hub and handheld virtual pet, including circuit boards, enclosures and firmware. Engineers at Diode Computers physically built and tested them. The companies report working prototypes without circuitry fixes.
Our interpretation. Editorial interpretation: coordinating geometry, component constraints, software and manufacturing moves beyond isolated coding tasks. Working hardware supplies a concrete check, while human fabrication and testing remain part of the system.
Main limitation. Lab-and-partner demonstration, not independent evaluation or a broad success-rate estimate. Anthropic has a commercial relationship with Diode. The three prototypes have not undergone electrical validation, antenna tuning, thermal testing or certification, and were not designed for mass production. Minimal prompting still included human directives; this was not autonomous physical manufacturing.
Next evidence. Independent builds, complete attempt and intervention logs, measured costs, durability and certification tests, and more complex designs that learn from instrument measurements.