Saved daily version · 2026-10-29
Daily evidence notes
Google announced Gemini 4 Carbon with stronger reported cyber and long-horizon performance and restricted initial access. Newly reviewed reports describe Anthropic’s real-software exploit chains and Axiom’s large-scale formalization of a major mathematical theorem. These are developer-reported results, not independent confirmations; the central tests are whether capabilities reproduce, safeguards hold, and formal statements faithfully capture the intended mathematics.
Sources checked and limits of this review
Reviewed developments since the October 28 briefing, including newly reviewed earlier work discovered through reporting leads. Checked METR updates, ARC Prize’s blog, Epoch AI’s latest work, Google DeepMind news, OpenAI research/news and Anthropic research, plus independent research leads. Carbon’s October 28 announcement is new in this window. Anthropic’s October 27 cyber report and Axiom’s October 26 formalization report retain their original dates and are newly reviewed, not newly published. Opened the linked Chrome vendor release notes and Axiom proof portal; neither establishes independent reproduction of the complete model workflow in this review. No comparably material new primary finding was verified from METR, ARC Prize, Epoch AI or OpenAI. Coverage is bounded, not exhaustive. SI relevance is editorial interpretation. No overall wisdom-trajectory verdict, proceed/stop recommendation or human approval is claimed. Source opening failed; saved README rules were read from the previously verified version-25 checkout. Routine publication uses the persistent updater, not a rebuild.
Evidence note 1 · Lab-reported capability and safety results; restricted rollout
Gemini 4 Carbon raises reported cyber performance; access starts with selected defenders
Google announced Carbon on October 28, reporting 88.2% on vulnerability detection in SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench patching. These capability tests ran without guardrails. Google also reports internal workflows lasting days and fewer high-severity misalignment alerts than Argon. Initial access is restricted to selected cyber defenders and trusted partners while the lab gathers feedback and participates in voluntary US pre-release testing.
Our interpretation. Editorial interpretation: connecting vulnerability discovery, exploit validation and repair could make sophisticated security work more broadly available. The same combination creates misuse potential. Progress toward wise SI depends on whether deployment boundaries and institutional response keep pace—not on treating stronger scores or a defensive launch label as proof of safety.
Main limitation. These are lab-reported comparisons, not independent replication here. External benchmark names do not make a lab-run evaluation independent. Internal alert reductions depend on what the monitor detects, and days of activity do not establish days of reliable progress. Finite tests with zero observed failures cannot establish general immunity.
Next evidence. Watch for independent equal-budget evaluations, public pre-release findings and access terms, realistic long-horizon success and failure rates, and tests of the deployed safeguard stack under sustained adaptive attack.
Evidence note 2 · Lab-reported exploit case studies; newly reviewed; no independent workflow replication
Newly reviewed: Anthropic describes models chaining flaws across real software defenses
Anthropic’s October 27 report describes five attack chains produced with an unreleased research model after initial prompting, spanning browser, phone and operating-system targets. The lab says the vulnerabilities were reported to vendors and fixed. Its workflow used parallel scanning agents, diverse search strategies and a verification pipeline. The report concerns work over preceding months, not five new incidents occurring today.
Our interpretation. Editorial interpretation: finding isolated bugs is different from combining weaknesses to cross several security boundaries. Such composition is relevant to powerful agents acting in environments with imperfect containment. Whether this produces less harm depends on disclosure, patch quality and adoption, not merely how many flaws a model finds. A security layer’s intended purpose must remain open to testing against its actual interactions.
Main limitation. The examples were selected by the developer and are not a measured success rate over all attempted targets. Model identity, total compute and the full human contribution are not quantified sufficiently here to establish reproducibility or autonomy. Vendor patches support remediation of particular flaws, not the entire claim about the model workflow.
Next evidence. Look for independently reproduced, authorized evaluations with budgets and failed attempts disclosed; vendor-confirmed patches and adoption; and comparisons of defensive repair speed with the time and cost required to develop attacks.
Evidence note 3 · Developer-reported formalization; newly reviewed; independent audit pending
Newly reviewed: Axiom says agents formalized the classification of finite simple groups
Axiom’s October 26 report claims a complete Lean formalization of the Classification of Finite Simple Groups, which lists the basic building blocks of finite symmetry. It reports roughly 43.1 million lines of code produced over 24 days by AI agents directed by engineers and mathematicians. The project built on existing mathematics and formalizations, repaired intermediate arguments and published a proof portal identifying the final theorem and checking instructions.
Our interpretation. Editorial interpretation: this would be evidence of coordinating a large, dependency-heavy research task, not of discovering the classification from scratch. The revealing difficulty is keeping each checked statement connected to the intended final claim. Formal verification can expose missing assumptions, but a green build is not a substitute for checking that the specification says what readers think it says.
Main limitation. This is the developer’s account, newly reviewed here; we did not rebuild the proof or audit its definitions, assumptions or certificates. Human mathematicians and engineers were integral. Code volume and elapsed time do not isolate the model’s contribution or establish general research autonomy.
Next evidence. Look for independent builds and audits of the final statement, its axioms and dependency closure; specialist review of repairs; and disclosed compute, human effort and failure history. Test transfer to other long research projects.