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

Saved daily version · 2026-10-13

Daily evidence notes

Three October 12 releases materially extend the record. Meta launched Muse Wonder and Muse Wonder Full for agentic work, but its broad capability claims mix internal, provider-reported, and third-party evaluations and await common-harness independent testing. Anthropic reported a contained experiment in which a large model could propagate a smaller inference payload through a deliberately easy emulated network; its forecast that small self-hosted replicators may arrive by year-end is explicitly extrapolative. Anthropic also reported a 90–99% fall in reasoning tokens extracted by suspected distillers after new controls, while acknowledging that durability against sophisticated adaptive attacks is unknown. METR, ARC Prize, Epoch AI, OpenAI, and Google DeepMind indexes showed no newer material publication; recent arXiv listings were also checked. Editorial interpretation: agent capability, uncontrolled persistence, and capability diffusion are advancing as linked technical and institutional questions, but these provider reports do not establish general intelligence, real-world prevalence, or an overall trajectory toward wise superintelligence.

Evidence note 1 · Lab-reported model release and capability/safety evaluation

Meta releases Muse Wonder agent models with broad long-horizon claims

Meta released Muse Wonder through its Model API and Muse Code, and Muse Wonder Full to Muse Code subscribers. The model card lists a 1-million-token context and reports evaluations spanning professional agents, memory, long-running software work, coding, health, finance, and law. Meta describes “Active Memory” and “Active Focus” as the ability to retrieve relevant prior context, persist through blockers, and check before important actions.

Our interpretation. Agents that can remember personal context, use tools, and sustain work across long tasks are more consequential than isolated question-answering gains because capability depends on the whole deployed system—model, memory, tools, permissions, and oversight. This is a meaningful product and evaluation milestone, not evidence of superintelligence.

Main limitation. LAB CLAIM: The scorecard blends Meta-run results, provider-reported results, and third-party leaderboards; for some comparisons it uses the highest internally validated, self-reported, or official score. Several headline constructs and benchmarks are internal, and competitor harnesses may not be equally optimized. The model card says testing cannot cover all scenarios and recommends human confirmation for irreversible actions. Meta’s reported 0% harmful-action rate on one agentic-misalignment evaluation does not prove absence of loss-of-control risk.

Next evidence. Independent evaluations under a common harness; task-level results on unseen long-horizon work; rates of memory error, privacy leakage, unsafe tool use, and recovery from bad plans; and a fuller safety report with external review and uncertainty estimates.

Evidence note 2 · Lab-reported frontier red-team experiments plus an explicit capability forecast

Contained tests narrow the gap to autonomous replication on cheap devices

Anthropic’s Frontier Red Team reports that a 9B model set up inference on 8 of 9 tested hardware types, and that a 26B model ran at about 0.39 tokens per second across 15 low-cost Android TV boxes. In a firewalled eight-node Android-emulator fleet with an unauthenticated port, GLM-5.2 propagated a Llama-3.1-8B inference payload to all eight nodes in four of five trials and seven nodes in the fifth.

Our interpretation. Autonomous replication matters because an agent able to acquire distributed compute and preserve itself could become harder to contain. The result usefully separates inference setup from exploitation and propagation, showing progress on a practical bottleneck rather than treating replication as a single vague capability.

Main limitation. LAB CLAIM AND FORECAST: This was not a self-hosted model copying itself across real compromised devices. The capable GLM-5.2 controller ran on separate infrastructure, the propagated payload was a smaller model, and network access was intentionally trivial—making the result an upper bound on real-world success. The prediction that small models may cross the capability threshold by the end of 2026 extrapolates a general capability index that is only a proxy for replication ability.

Next evidence. Independent reproduction; a small model running locally and copying its own weights and harness; tests against authenticated, patched, heterogeneous devices; realistic measures of detection and persistence; and demonstrations that ISP or device-level defenses can drive onward replication below replacement.

Evidence note 3 · Lab-reported security intervention and operational measurement

Anthropic reports sharply lower reasoning extraction after anti-distillation controls

Anthropic says new controls—restricted movement of preserved reasoning, a two-stage extraction classifier, account limits, and bans—reduced reasoning tokens harvested by suspected distillers on its platform by roughly 90–99%, or 10–100 times. It reports a classifier false-positive rate below 0.1% and says tens of thousands of associated accounts were banned.

Our interpretation. Distillation can spread frontier capabilities faster and more cheaply than independent training. Whether safeguards, evaluation standards, and governance travel with those capabilities affects the wisdom of the resulting ecosystem; restricting extraction may slow diffusion, but it also concentrates control and creates trade-offs for legitimate users.

Main limitation. LAB CLAIM: The extraction metric, attacker classification, baseline, and intervention effects are reported by the provider operating the platform and have not been independently audited here. Multiple defenses changed together, so their individual contributions are unclear. Anthropic says it has not yet faced a highly sophisticated, persistent circumvention attempt, expects an ongoing adaptive contest, and acknowledges integration costs, blocked benign requests, and charges on some blocked traffic.

Next evidence. Independent audits of detection accuracy and causal effect; red-team bypasses using new extraction strategies; durability over months; transparent accounting of legitimate-user harms; and evidence about whether blocked transcript access actually slows downstream capability replication.