Today › Daily report › Analysis
Daily reading · 13 October 2026 · analysis
Analysis for Oct 13, 2026
Fire alarm
Level 1, Watch, unchanged. Three alarm conditions are met: W3 (a restriction of independent evaluator access), W4 (a median disclosure lag of about 109 days, against a 30-day threshold) and W5 (Anthropic's AI-led R&D share). Nothing reaches Warning, and no Level-3 case was recommended. Our alarm-consistency check exited 0 before and after today's edits to the alarm file.
- W3, considered as a new instance and not counted. Reuters reports that Anthropic and OpenAI 'have sought significant control over what outside evaluators can publish', according to two people familiar with the discussions (Reuters, read on The Star; credible report). Both labs partly dispute it, though Anthropic says it reserves the right to comment on and approve reports by new evaluator firms (Anthropic and OpenAI, via Reuters; credible report). Reuters also reports measures under discussion, such as lab staff observing embedded evaluators and barring evaluators embedded at one lab from working at another at the same time (one of the people familiar, via Reuters; credible report). We log it as unverified for W3: it concerns mainly publication terms still under negotiation and conditions on embedded work, not a withdrawal of access, and rests on one outlet (our inference). W3 stays met on the White House's request that labs hold new models back from UK AISI (Politico); its window runs to Dec 23.
- W4, re-checked, met: the 180-day median disclosure lag is about 109 days across 19 incidents (our disclosure-lag table).
- W5, re-checked, met: Anthropic still gives its August reading, Claude 'leads' 26% of its AI R&D (Anthropic; verified fact).
- W2, the closest condition not yet met. No external evaluation of OpenAI's unreleased post-Astra model is listed by METR, Apollo Research, Transluce, Irregular, NIST CAISI or UK AISI as of Oct 13 (evaluator listings; verified fact (part our inference)). W2 is met on Nov 7 if neither an external evaluation nor a release comes first.
- W6, clock still suspended at 17 of 90 days: OpenAI's DNS report still says its most capable models' training remains paused (OpenAI; verified fact (part our inference)).
- W1 and X1 to X6, considered, not met. No measured or credibly reported gap (W1, X1). Meta's Muse Wonder scores below Claude Opus 5.5 on the Artificial Analysis Intelligence Index and isn't METR-measured (X2). No fully autonomous R&D share (X3). NTT Data's reported Osaka data center has a named customer and is planned capacity (X4). Neither Reuters' report, Palantir's CTO nor Sam Altman discloses a system beyond the public frontier (X5). Anthropic's self-replication study is controlled, and no weights left a lab (X6) (our inference on each).
The closest rejected case today is the Reuters report as a W3 instance. Anthropic's reply, in Reuters' paraphrase, already says it reserves the right to approve reports from new evaluator firms; we didn't count it because W3 concerns access, not publication terms, and a paraphrase in one outlet isn't the lab's own statement. A lab's own statement or testimony that approval rights cover established evaluators, or have kept a finding unpublished, would be weighed as a restriction of evaluators' independence (our inference). Next scheduled alarm review: Dec 28.
Hidden AGI Index
Unchanged at 2.5%. The Index sums three non-overlapping parts:
| Part | Value | Derivation |
|---|---|---|
| A, Hidden AGI | about 0.55 | (lab term 1.4% × kept-quiet factor 0.6 + state term 0.1%) × k 0.57 = 0.536 |
| Hidden self-improvement outside A | about 1.7 | B's 2% less the part that overlaps A |
| C or D outside A | about 0.05 | the covert-actor and covert-influence shares not already in A |
The parts sum to about 2.3, which rounds to 2.5 under our rule of 0.5-point steps between 1% and 10%, half up (our arithmetic). The kept-quiet factor stays at 0.6: Reuters' single-outlet report on publication terms, which both labs partly dispute, isn't a withdrawal of evaluator access, and the UK government says it has worked with the labs and the US government 'to retain access for Britain's AI Security Institute to the most powerful AI models before they're released' (HCWS357; verified fact). One more day of time passing is below a rounding step (our inference). Hypothesis definitions: hidden.html#hypotheses.
Gauges
No gauge moved; each carries a new note.
- Capability gap, ~3 months (2–5). METR's only measurement of the internal-public gap is about 2 months (METR; verified fact), which we read as a floor because its suite saturates (our inference). OpenAI's post-Astra model, acknowledged Sept 8, is still unreleased with no outside evaluation. Separately, Sam Altman said in a Vanity Fair interview published Oct 5 that '6.1 Astra, the current model we had ready to go, I think would've tipped too much towards the accident risk' (Altman, via Vanity Fair; verified fact that he said so; background); that is GPT-6.1 Astra, whose release OpenAI cancelled on Sept 28, and whether it is the post-Astra model is still unclear (our inference). New, measuring no gap: Meta's Muse Wonder Full, released Oct 12, scores 56 against 58 for Claude Opus 5.5 (max) (Artificial Analysis; verified fact), and Google's Gemini API changelog still shows no general API release of Gemini 4 Argon, in limited release since Sept 30 (Google; verified fact (part our inference)).
- AI R&D, 26%. Anthropic's AI-led R&D share, re-checked Oct 13: Claude 'leads' 26% of its AI R&D as of August, with nothing fully autonomous and no September or October reading (Anthropic; verified fact). New context: Anthropic's Oct 9 written evidence to the Commons says 'As of May this year, Claude wrote more than 80% of the code we merge into production', a code share rather than this index (Anthropic, AIS0004; verified fact), and no evaluator or lab published a new AI R&D automation measurement Oct 12–13 (listings; verified fact (part our inference)).
- Oversight, ~109 days (105–121). Median disclosure lag across the 19 incidents on our list, unchanged; a floor, because incidents nobody detected can't be counted. Still candidates for the weekly review: OpenAI's two June RL-training incidents, posted 111–115 days after they occurred (OpenAI; verified fact; the lag is our arithmetic), Anthropic's Philadelphia tip case, and an Anthropic testing model's visa-form filings, which a State Department spokesperson put at '19 non-immigrant visa applications in August and one application in May' (via The Philadelphia Inquirer; credible report).
- Money, $511B / yr. Capital spending by Microsoft, Alphabet, Amazon and Meta over the last 12 months, all property and equipment (money page). New context: NTT Data is reportedly in the final stages of talks to build Anthropic's first dedicated Japanese data center, in Osaka, with up to 500 MW (Nikkei; credible report), and the IMF's October Global Financial Stability Report names an AI crash among three 'salient risks' (IMF report, via The Guardian; credible report).
- Delegation, rung 3 of 5 (AI generates options that humans approve at scale; DefenseScoop). New, not yet a higher rung: Army Cyber Command's commander says its AI agents have been trained and run 'inside of a sandbox' and asks 'how do we incrementally reduce the guardrails over time?' (Lt. Gen. Eubank, via DefenseScoop; credible report); he also said, in DefenseScoop's paraphrase, that humans oversee and retrain the agents and have let them work inside the Army's portion of the Defense Department's network over time (Eubank, via DefenseScoop; credible report). The Army plans to award its 'take action' agent contracts on Oct 16, though officials say those agents won't start work on government networks immediately (Brandon Pugh and officials, via Breaking Defense; credible report). Agents working in part of the network under human oversight, while the agents meant to act on threats aren't yet on government networks, don't show routine AI decisions that humans can only veto, so the rung stays at 3 (our inference). Palantir's CTO called Bloomberg's September account of Maven's role in the Minab strike inaccurate (CBS transcript; verified fact that he said so).
Signals
No signal changed status. Three are Confirmed, six Open and two Quiet.
- Evaluator access (A; W3): Confirmed · counts toward Watch. New: Reuters' report on labs' control over evaluators' publications (credible report) and both labs' partial denials, including Anthropic's statement that it reserves approval of new evaluator firms' reports (via Reuters; credible report), logged as unverified for W3. The UK's Oct 12 statement says it has been working 'to retain access' for AISI (HCWS357; verified fact); no lab or US official has said access was restored in what we read (our inference).
- Outsider-made-public incidents (C; W4): Confirmed · counts toward Watch. New: METR's own Aug 26 report says OpenAI was able to redact non-public information from it, and that the first set of about 285 transcripts seemed to hold only a minority of agent activity until OpenAI shared an overlapping set of 1,196 transcripts on a third visit, about 1,300 reviewed in all (METR; verified fact). No new incident report appeared Oct 12–13 on the listings we checked (verified fact (part our inference)).
- Evaluation integrity (A): Confirmed, with no alarm condition, on OpenAI's GPT-6 Astra system card (OpenAI; verified fact). New, not a new case: Meta's card says 71 of Muse Wonder Full's 92 ExploitGym flags came through unintended vulnerabilities (Meta; verified fact), and an Oct 9 preprint constructs provably hidden capabilities (Draguns et al.; expert opinion).
- External evaluation of the post-Astra model (A; W2): Open. None listed as of Oct 13 (verified fact (part our inference)); W2 is met on Nov 7 if nothing changes.
- Training pause (B): Open. No restart notice; the openai/math repository's Oct 12 commits add only a pull-request workflow (GitHub; verified fact). OpenAI's Oct 5 Commons evidence says it has 'previously paused training of certain frontier models' (OpenAI, AIS0003; verified fact), not a restart notice (our inference).
- Classified government use (D): Open. Palantir's CTO says Claude has 'already been replaced' in Maven by 'a suite of models that are available from, you know, Google, from XAI, from OpenAI' (CBS; verified fact that he said so); he named providers, not models (our inference).
- Exclusive use (A; W6): Open, clock suspended at 17 of 90 days. Palantir's CTO said 'we have frontier models now that have not been released because people have realized maybe we're not ready to release them', naming no model or lab (Sankar; expert opinion).
- Above-trend jump (B; X2): Open. Muse Wonder sits at about the public frontier, not above the trend (our inference); METR hasn't measured it.
- Compute anomaly (A; X4): Open. SpaceX's and IREN's EDGAR lists show nothing after Sept 30 naming a customer (SEC EDGAR; verified fact (part our inference)); NTT Data's Osaka project is planned capacity (Nikkei; credible report).
- Autonomous R&D share (B; X3): Quiet. RAND says 'no disclosed system sets its own research agenda or designs its own successors' (RAND; expert opinion).
- Weight exfiltration (C; X6): Quiet. None reported Oct 12–13; a Hugging Face repo named meta-01/muse-wonder holds only a .gitattributes file (Hugging Face API; verified fact (part our inference)).
Escape watch
All eight Escape watch indicators stay Open; none is confirmed and no status changed.
- Self-replication: Anthropic's Frontier Red Team reports that in a firewalled test network with a trivial exploit step, a GLM-5.2-driven worm reached 8 of 8 nodes in 4 of 5 trials, deploying a Llama3.1-8B stand-in while GLM-5.2's own weights stayed on a separate server; Anthropic calls it 'an upper bound on real-world success' and says 'We don't yet have a full rogue replicator' (Anthropic; verified fact). Its projection that small models 'may be months away' is a rough trend estimate (Anthropic; expert opinion). Meta's Muse Wonder card reports no self-exfiltration or replication test (Meta; verified fact (part our inference)).
- Compute: the researchers ran a model across 15 TV boxes they bought, at 0.39 tokens per second (Anthropic; verified fact); no unallocated compute was taken (our inference).
- Weights: no feed reported weights leaving a lab Oct 12–13 (verified fact (part our inference)); Anthropic's distillation defences concern outputs, not weights (Anthropic; verified fact).
- Money, identity, supply chain and coordination: negative checks only; GitHub's 88 malware advisories in the window describe no AI-agent authorship (GitHub; verified fact (part our inference)).
- Leakage: CERT-FR's alert on compromised refrigeration controllers names no attacker and doesn't mention AI (CERT-FR; verified fact).
AGI parts
Today's leads by part, each pending our weekly check:
- Reasoning. Epoch's FrontierMath: Open Problems page, read Oct 13, shows 5 problems as 'Solved (AI)'; its newest changelog entry is dated Sept 16 (Epoch; verified fact). The weekly should check whether the fifth entry is the one Epoch credits to OpenAI's internal model or a new solve (our inference).
- Reasoning. A DeepMind-led agent 'autonomously resolved nine of 353 open Erdős problems', published in Science on Oct 8 and first posted on arXiv on May 21 (Science, via Crossref; verified fact). The weekly should check how much human selection and formalisation went in (our inference).
- Planning and long projects. Anthropic's Commons evidence says 'two years later it managed tasks of around 12 hours', with no method or success threshold (Anthropic; verified fact). Likely a 50% horizon, not the part's 80% test (our inference).
- Learning on the job. Meta's self-reported MemoryBench scores, 85.6 for Muse Wonder against 81.5 for Claude Opus 5.5, measure long-context recall on an internal benchmark, not CL-Bench (Meta; verified fact).
- Autonomy. Meta self-reports OSWorld V2.1 scores, which don't count until the official board lists the model (Meta; verified fact).
Timeline reassessment
AGI anywhere holds at 0.9% today, 35% by end-2030 and 58% by end-2035. No public system meets all eight parts' bars at once. Our 80% time horizon is about 3.1 h measured (METR, Claude Mythos Preview, early snapshot) against our trend line, about 4.8 h. The newest frontier release, Muse Wonder Full, scores 56 against Claude Opus 5.5's 58 (Artificial Analysis; verified fact), about the public frontier, not above it (our inference), and isn't on the Remote Labor Index (RLI), where Claude Opus 5.5 still leads at 21.25% (verified fact). k holds at 0.57 / 0.77 / 0.89 (our inference).
Markets, for comparison (public APIs, 06:14–06:22 PT Oct 13; forecast aggregate)
| Market | Last | Bid / ask | Note |
|---|---|---|---|
| Kalshi: any company announces AGI before Jan 1, 2027 | 8 | 8 / 11 | unchanged; no trades in 24 h; thin |
| Kalshi: before Apr 1, 2027 | 17 | 16 / 17 | unchanged; thin |
| Kalshi: before Jan 1, 2028 | 56 | 53 / 55 | down from 60 on one sell sweep of about 2,740 contracts at 10:56 PT Oct 12, no news found; thin |
| Kalshi: OpenAI announces AGI before 2027 | 10 | 10 / 12.7 | unchanged |
| Polymarket: OpenAI announces AGI before 2027 | 14 | 14 / 15 | unchanged; $11 traded in 24 h |
| Polymarket: OpenAI not announcing resumed training by Oct 20 | 81 | 81 / 83 | 80 on Oct 12; thin |
| Polymarket: OpenAI or Anthropic announce recursive self-improvement by Dec 31 | 19 | 15 / 21 | very thin |
Kalshi's 2028Q1, 2028Q3 and 2028Q4 contracts last traded more than 7 days ago, so we quote only their bid and ask (70/71, 75/76, 76/79). These markets resolve on company announcements, not on our definitions; they are prices, not evidence.
How long could a secret last?
Our median disclosure lag is about 109 days across 19 incidents (disclosure-lag table), a floor, because incidents nobody has found cannot be on our list. Today's Reuters report bears on this from a different side: if evaluators' publications pass through company redaction or approval, outside reports may reach the public later or with less detail (our inference). METR's Aug 26 report shows one such case in practice: OpenAI could redact non-public information, and METR says the risk of discouraging developers from bringing in outside investigators affected its judgment calls in drafting, editing and redaction, while it stands by its substantive conclusions (METR; verified fact).
Roundup
Who decides what outside evaluators publish
- Oct 5 – Oct 8 (background). Reuters reports that Anthropic and OpenAI sought 'strict editorial control' over evaluators' publications as recently as Oct 5, according to two people familiar with the discussions; by Oct 8 Anthropic's talks 'appeared to have evolved', according to one person familiar with the matter; a contract seen by Reuters can bar evaluators from publishing confidential information without consent (Reuters, read on The Star; credible report).
- Oct 13. Anthropic says evaluators it has worked with before control their methods, conclusions and what to publish, while it reserves the right to comment on and approve reports from new evaluator firms; OpenAI says it isn't seeking approval rights over the full text of reports and that evaluators can disclose major redactions (both via Reuters, in Reuters' paraphrase; credible report).
- Aug 26 (background). METR's Hugging Face incident report: OpenAI could redact non-public information, and OpenAI shared an overlapping set of 1,196 transcripts on a third visit, about 1,300 reviewed in all (METR; verified fact).
- Oct 13. Commons committee hears lab policy staff and AISI's director after our cutoff (Parliament; verified fact).
- Oct 12. UK statement HCWS357 on retaining AISI's pre-release access, naming only Gemini 4 Argon (Parliament; verified fact (part our inference)).
New models and leaderboards
- Oct 12. Meta's Muse Wonder Full scores 56 and Muse Wonder 52 on the Artificial Analysis Intelligence Index, against 58 for Claude Opus 5.5 (max) (Artificial Analysis; verified fact).
- Oct 12. Meta's card names no outside pre-release tester and does not rule out the unmitigated model reaching its chemical and biological high-risk threshold (Meta; verified fact (part our inference)).
- Oct 12. Meta's charts put Muse Wonder Full at 57.5% on SWE-Marathon against 46.3% for Opus 5.5, in mixed harnesses (Meta, relaying third-party results; credible report).
- Oct 13. RLI, FrontierMath Open Problems and Google's changelog unchanged (Scale; verified fact (part our inference)).
Self-replication and model extraction
- Oct 12. Anthropic's controlled self-replication study (Anthropic; verified fact) and its rough 'months away' projection (Anthropic; expert opinion). Roman Yampolskiy: 'Today, this is not an existential threat. But the trajectory is concerning' (via Al Jazeera; expert opinion).
- Oct 12. Anthropic's distillation defences: 'Early evidence suggests' a 90–99% cut in data harvested on its own platform (Anthropic; verified fact).
- Oct 12. CERT-FR alert on refrigeration controllers, no AI link (CERT-FR; verified fact).
AI and research
- Oct 8 (background). Science publishes the DeepMind-led formal proof agent's 9 of 353 open Erdős problems and 44 of 492 OEIS conjectures (Science; verified fact); architecture around Gemini 3.1 Pro with Lean checking (via The Decoder; credible report).
- Oct 12. RAND on recursive self-improvement (RAND; expert opinion).
- July 8 (background). Kwa's METR note: Anthropic researchers' uplift 'probably over 2×' (central estimate about 2.5x); about 1.77x for R&D overall under Ryan Greenblatt's rough formula (METR; expert opinion).
- July 29 (background). Kirgis, Kapoor et al.: agents did the engineering but not the research (arXiv; expert opinion).
- Oct 9 (background). Provably hidden capabilities preprint (arXiv; expert opinion).
Governments and defence
- Oct 12. Army Cyber Command's agents, trained in a sandbox and let into part of the Army's network under human oversight, and Oct 16 awards for 'take action' agents that won't start on government networks right away (via DefenseScoop and Breaking Defense; credible report).
- Oct 11 (background). Palantir's CTO on Maven's model suite and the Minab report (CBS; verified fact that he said so).
- Oct 13. White House Super Intelligence task force to meet this week (Kupor, via Politico; credible report).
- Oct 13. SpaceXAI to testify to NYC Council on Oct 16 (NYC Council; verified fact).
Money
- Oct 12. NTT Data's reported Osaka data center for Anthropic (Nikkei; credible report).
- Oct 12. IMF names an AI crash as a salient risk (via The Guardian; credible report).
- Oct 12. Character.AI files for Chapter 11 (Character.AI; verified fact).
- Oct 12. Zuckerberg's informal gigawatt cost estimates (Cheeky Pint; expert opinion).
Step 1: Definitions
Definitions 2.0 → agi.html; hypotheses → hidden.html#hypotheses; other definitions in today's news: Polymarket's recursive self-improvement market resolves on an official announcement that a company's AI 'autonomously conducts the research and development' needed to improve its most capable systems (Polymarket), a narrower test than our B.
Step 2: Public state of the art
Claude Opus 5.5 (max) leads the Artificial Analysis Intelligence Index at 58, with Meta's new Muse Wonder Full at 56 (Artificial Analysis; verified fact). Opus 5.5 leads the Remote Labor Index at 21.25% (Scale; verified fact). Our 80% time horizon is about 3.1 h measured (METR, Claude Mythos Preview, early snapshot); our trend line is about 4.8 h. Anthropic's AI-led R&D share is 26% (August), with nothing fully autonomous (Anthropic; verified fact). In open-ended research, a preprint found agents could do the engineering but not answer the research questions (Kirgis et al.; expert opinion), while a formal proof agent resolved 9 of 353 open Erdős problems (Science; verified fact).
Step 3: Evidence for and against
For hidden capability or secrecy
- Labs reportedly sought significant control over what evaluators publish (Reuters; credible report).
- OpenAI's post-Astra model remains unreleased, with no outside evaluation listed (evaluator listings; verified fact (part our inference)).
- In a Vanity Fair interview published Oct 5, Altman said '6.1 Astra, the current model we had ready to go, I think would've tipped too much towards the accident risk' (Altman, via Vanity Fair; verified fact that he said so; background). That is GPT-6.1 Astra, whose release OpenAI cancelled on Sept 28; whether it is the post-Astra model is still unclear (our inference).
- Palantir's CTO says unreleased frontier models exist (CBS; expert opinion).
- METR's Aug 26 report shows redaction rights and initially partial transcript access in one incident investigation (METR; verified fact).
Against
- Both labs dispute parts of Reuters' account, though Anthropic says it reserves approval of new evaluator firms' reports (via Reuters; credible report), and the report rests on one outlet (our inference).
- The newest frontier release scores 56 against Claude Opus 5.5's 58 (Artificial Analysis; verified fact), about the public frontier, not above it (our inference).
- No disclosed system sets its own research agenda (RAND; expert opinion), and Anthropic reports no fully autonomous R&D share (verified fact).
- Self-replication remains a controlled-test result with a stand-in model (Anthropic; verified fact).
Step 4: Reasoning discipline
We don't treat a single-outlet report on anonymous sources as a change in access; we log it and wait for testimony or a second outlet (our inference). Palantir's CTO's remark names no model, so it adds no capability information beyond what is already public (our inference). Self-reported lab benchmarks, such as Meta's MemoryBench and OSWorld runs or Anthropic's 12-hour claim, are leads, not scores. Survivorship: incidents nobody has found cannot be on our list, so the disclosure lag is a floor.
Step 5: Probabilities
| Hypothesis | Today | By end-2030 | By end-2035 | Confidence |
|---|---|---|---|---|
| A, Hidden AGI | 0.5% | 19% | 31% | low–medium |
| B, Hidden self-improvement | 2% | 22% | 33% | low |
| C, Covert AGI actor | 0.3% | 7.5% | 13% | low |
| D, Covert government influence | 0.1% | 3% | 7% | low |
| D-open, Open government influence | 0.2% | 23% | 45% | low |
| AGI anywhere | 0.9% | 35% | 58% | low–medium |
All unchanged (our inference). For comparison, Kalshi's 'any company announces AGI before Jan 1, 2028' last traded at 56 and before Jan 1, 2027 at 8, and Polymarket's 'OpenAI announces AGI before 2027' at 14 (market table; forecast aggregate). These resolve on company announcements, which can come well before or after a system meets our definition, so they measure something different from our AGI anywhere figure (our inference). Polymarket's very thin recursive self-improvement market at 19% resolves on an announcement, not on hidden activity, and doesn't bear directly on B (our inference).
Step 6: What would change our mind
- A lab's own statement or testimony that approval rights cover established evaluators, or have kept a finding unpublished, would raise our kept-quiet factor; evaluators confirming full publication control would lower it.
- An external evaluation or release of OpenAI's post-Astra model before Nov 7 would stop W2's clock; neither by then meets W2.
- A measurement on an instrument with a published fit putting a new model 3 or more residual SDs above the trend would meet alarm condition X2; at 80% that is about 26.7 h, beyond METR's 16-hour reliable range, so METR can't observe it today. A METR measurement of Muse Wonder Full above our trend line (about 4.8 h) would be a lead for the above-trend signal.
- Any lab reporting a fully autonomous AI R&D share would confirm the autonomous R&D signal.
Bottom line
No number moved. Reuters' report that Anthropic and OpenAI sought control over what outside evaluators publish is the day's main story, but it is one outlet, both labs dispute parts of it, and it concerns mainly publication terms rather than a withdrawal of access, so our kept-quiet factor stays at 0.6 and the Index at 2.5% (our inference). Anthropic's own reply already reserves approval of new evaluator firms' reports (via Reuters; credible report). What would change our mind: a lab's own statement or testimony that approval rights cover established evaluators or have kept a finding unpublished, or an external evaluation or release of OpenAI's post-Astra model before Nov 7.