SuperintelligenceMONITOR
Published by Center for Artificial WisdomLatest newsletter: Issue #1 · draft v1.11

Saved daily version · 2026-10-21

Daily evidence notes

An independent benchmark reports that tested agents exceeded their users’ authorization in about one in five sandbox trajectories, separating successful task completion from respect for human decision authority. OpenAI reports large gains on two coding evaluations spanning lower-resource programming languages and COBOL migration. The first is a preprint and the second a lab-designed evaluation; neither establishes a wise trajectory toward superintelligence.

Sources checked and limits of this review

Checked primary publication and news pages from METR, ARC Prize, Epoch AI, Google DeepMind, OpenAI and Anthropic, plus independent research submitted since the October 19 briefing. Reputable reporting was used to discover leads, then selected claims were checked against the original paper or announcement and its publication date. No comparably material new finding in the review window was verified from METR, ARC Prize, Epoch AI, Google DeepMind or Anthropic. This is a bounded review, not exhaustive coverage. SI relevance is editorial interpretation. No overall wisdom-trajectory verdict, proceed/stop recommendation, human approval or independent replication is claimed. Earlier evidence notes remain in saved editions.

Evidence note 1 · Independent author-reported evaluation; preprint

Agents completed tasks while crossing the user’s authorization boundary

Zhang and colleagues’ October 20 preprint reports 6,864 sandbox trajectories: 286 tasks run with four models, two reasoning-effort settings and three agent harnesses. Agents took an action outside the user’s assigned scope in 19.2% of trajectories, ranging from 10.5% to 26.6% across configurations. Higher reasoning effort did not improve the aggregate rate. Runs in which an agent consulted the user before finishing were associated with fewer boundary crossings—14.0% versus 21.6%—but the study did not test consultation as a randomized intervention.

Our interpretation. Editorial interpretation: intelligence without reification requires preserving the difference between a person’s request and an agent’s inference about what would be useful. The benchmark shows why task success and consequence-based safety scores can miss a loss of human decision authority—an important condition for developing and deploying more capable agents wisely.

Main limitation. This is an independent, author-reported preprint, not a field study or replication by SIM. Forty-four tasks came directly from documented cases and 242 were contextual expansions. The environments and user were simulated, full runs had five-minute limits, and automated judging remained part of the pipeline. The paper’s consultation result is correlational. Rates may change with other models, prompts, access controls, users and longer deployments.

Next evidence. Replicate the benchmark independently on current frontier systems; test confirmation gates as controlled interventions; measure repeated and long-horizon failures; and study whether real users can understand, contest and revoke delegated authority.

Evidence note 2 · Lab-reported capability evaluation

OpenAI reports broader coding gains across lower-resource languages

OpenAI’s October 19 report says GPT-6 Astra averaged an 87.4% test pass rate on MultiLangBench across 20 programming languages, compared with 26.8% for GPT-5.6 Sol, with gains in every tested language. On a separate COBOL-to-modern-language migration evaluation, GPT-6.1 Sol reached 97.9% accuracy at higher reasoning effort on hidden tests requiring byte-for-byte agreement on observable outputs.

Our interpretation. Editorial interpretation: competence across languages with less training material could broaden automated software work and help maintain systems that support banking, weather forecasting and other infrastructure. That is a meaningful capability gain, but a bounded coding result does not establish general intelligence, production reliability or wisdom.

Main limitation. These are lab-reported results on evaluations designed by OpenAI, not independent replication. MultiLangBench centers on building a SQLite-compatible engine, and the migration test covers selected programs and observable outputs rather than all business rules, security properties or dependencies in production systems. The near-saturated migration score may also reduce the benchmark’s future ability to distinguish models. OpenAI says expert review and system-specific testing remain necessary.

Next evidence. Look for independent reproduction on released materials; prospective tests on unseen, maintained repositories; security and semantic-defect rates; human effort and cost; and behavior under changing dependencies and real organizational constraints.