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

Saved daily version · 2026-10-12

Daily evidence notes

Quiet-day review: no material new developments since the October 11 briefing were identified in the sources reviewed this morning. We checked METR's research, notes, and updates; ARC Prize; Epoch AI; Google DeepMind; OpenAI; Anthropic; and recent arXiv AI/ML listings. Recent reporting led back to announcements already available before the previous review. Epoch's benchmark hub bears an October 12 update date, but its latest described evaluation activity remains October 1; a refreshed page date alone does not establish a new result. Additional papers in the October 12 arXiv listings carry October 8–9 submission dates and are not presented as overnight breakthroughs. The three items below are retained from the October 11 edition, with their original October 9 source dates and limitations unchanged. They concern protein design, deliberately hidden capabilities, and failures across connected tools. Editorial interpretation: no new evidence selected today justifies changing those assessments or making an overall judgment about the trajectory toward wise superintelligence.

Evidence note 1 · Research-team reported model evaluation and laboratory experiment — preprint

Text-guided protein design reaches a limited laboratory test

Prompteina generates protein structures from written descriptions. Its authors report stronger results than baselines on 15 of 18 structural properties. In an agent-guided campaign, they generated 1,000 candidates, selected 90 for testing, and assayed 89; nine bound the GM2A target, including six with sub-micromolar binding. The paper was submitted October 9.

Our interpretation. A language interface can help researchers connect design intent to a specialized scientific model. The result concerns a combined workflow—an agent, design model, selection process, and human-run laboratory—not a general model independently mastering biology.

Main limitation. Most comparisons use predicted structure or interface confidence. The prospective laboratory test covers one target and selected candidates, with no matched experimental baseline campaign. Binding does not establish therapeutic benefit or safety. Apparent gains in prediction scores should not be read as laboratory success rates.

Next evidence. Independent replication, prospectively matched laboratory comparisons on more targets, functional tests, and whether iterative feedback improves measured outcomes rather than just prediction scores.

Evidence note 2 · Independent research — conditional theoretical result and proof-of-concept experiments

A constructed backdoor shows why failing to elicit a capability cannot prove its absence

Researchers deliberately built encrypted neural circuits that activate only with a secret trigger. Under stated cryptographic assumptions, recovering the hidden computation is hard even with access to the weights. Small proof-of-concept tests resisted several elicitation methods. The October 9 paper is listed as accepted at NeurIPS 2026.

Our interpretation. As capability audits help justify deployment, their conclusions must distinguish abilities demonstrated from abilities ruled out. Model provenance and checks for tampering matter alongside behavioral testing.

Main limitation. This studies deliberate weight manipulation, not evidence that current frontier models naturally conceal encrypted abilities. Experiments are small; full integration into a pretrained host remains future work. The construction is costly, and its presence may be detectable even when its trigger or payload is not. The proof does not show that all audits are useless.

Next evidence. Independent reproduction, tests of anomaly detection and training-provenance safeguards, realistic integration studies, and audit reports that state precisely which adversarial assumptions they exclude.

Evidence note 3 · Lab-reported agent red-team evaluation — simulated environment, preprint

Testing connected agents requires examining what tool actions do together

Meta researchers' X-AgentRed builds a graph of tool dependencies and uses execution feedback to search for harmful workflows arising from benign requests. Across four target models in a simulated ecosystem of 30 services and 914 tools, it uncovered more harmful paths and broader tool coverage than two baseline methods. The paper was submitted October 9.

Our interpretation. More capable assistants can act across services, where individually plausible steps combine into privacy or authorization failures. Evaluating the whole workflow helps keep human intent and permissions visible.

Main limitation. Each configuration was evaluated once in a fixed simulation. Model-generated service responses and model judges shape the results. These adversarially selected scenarios do not estimate real-world incident prevalence or establish rankings for today's frontier systems. The dependency graph also needs maintenance as tools change.

Next evidence. Independent replication in sandboxed real services, human validation of harms, repeated runs with uncertainty estimates, and defenses tested on unseen service combinations. Measure useful task completion alongside avoided unauthorized outcomes.