Today › Daily report › Analysis
Daily reading · 29 October 2026 · analysis
Analysis for Oct 29, 2026
Fire alarm
Fire alarm unchanged at Level 1, Watch (W3, W4, W5 met; alarm check consistent, exit 0, before and after today's edits; computed level 1, earliest drop Nov 28 if the rule stops holding). W2: day 51 of 60, met Nov 7 if OpenAI's post-Astra model gets no outside evaluation and isn't released first. W6: 45 of 90 under the main reading (Sept 8–24 and Oct 1 onward), met Dec 13. Dated rulings (2026-10-29) for W1–W6 and X1–X6. Nothing at Warning; no Level-3 case recommended; no review due (next Dec 28).
Conditions met today: W3, W4, W5 (Level 1, Watch). The closest rejected case is W6 and W2 for Google's Gemini 4 Carbon: Sundar Pichai said 'We are now deploying Carbon internally, and Carbon is a meaningful step up from Argon' (Pichai, via the Motley Fool and StockAnalysis transcripts; verified fact), while outside access starts with 'a subset of cyber defenders' (Google; verified fact). Google published Carbon's benchmarks the same day and says it is 'actively engaged in the U.S. government's voluntary process for pre-release model testing' (Google; verified fact), so the capability is acknowledged rather than kept for internal use, and neither clock starts (our inference). W1 is not met: a one-point lead for a disclosed model on the Artificial Analysis Intelligence Index is not a gap of 3 months or more (our inference). No case was referred to the tribunal. Nothing approaches Y1, Y2 or Y3: no source shows an undisclosed system meeting all eight parts, an AGI-level system acting covertly or shaping a government decision, or undisclosed work in which AI did most of the work on a successor with a 3x speed-up (our inference).
- W3: still met, window to Dec 23. Carbon's post names no UK AISI, while Google DeepMind's Argon report says 'AISI conducted evaluations of Gemini 4 Argon' (Google DeepMind; verified fact): an absence, not a reported restriction (our inference). The UK's extra £120 million for AISI (HM Treasury; verified fact) is funding, not restored access (our inference).
- W4: still met, ~115 days across 27 incidents (112–120), a floor. CrowdStrike's four campaigns and Dream's PHP2Shell operation are human-directed with public models and go to the weekly incident review, not today's median (our inference).
- W5: still met to Feb 15, 2027. Google's '35% more code with 30% fewer errors' is product code, not a new AI-led share of AI R&D (our inference).
- X3: not met. Pichai's expectation of 'continued acceleration because of RSI' gives no autonomous share or 3x estimate (Pichai, Google; expert opinion), and Google DeepMind's Argon report found 'no significant end-to-end acceleration of AI R&D' (Google DeepMind; verified fact).
- X4: not met. Microsoft's and Alphabet's Q3 spending and commitments, an unnamed Azure commitment, Samsung's capex, DOE's Constellation loan and Malaysia's server seizure are not an unattributed operating cluster (our inference).
- X1, X2, X5, X6: not met or can't be observed. No measurement or leak of a 6-month gap; Carbon's index and Blueprint-Bench 2 results sit on instruments with no published fit; no whistleblower or leaked document; no weight theft reported (our inference).
Hidden AGI Index
Unchanged at 2.5% (its parts sum to about 2.28). A is about 0.573, shown as 0.55 on the 0.05 grid: (lab term about 1.510% × kept-quiet factor 0.6 + state term 0.1%) × k 0.57. Hidden self-improvement outside A is about 1.66, shown as 1.65 (B's 2% less 0.6 × A). C or D outside A is about 0.05. The lab term rises from about 1.506% to about 1.510% for one day of time passing along our own forecast curve (the logistic path applied on Oct 24, about ×1.0026 a day). Gemini 4 Carbon, disclosed with its benchmarks published and one point ahead of the public frontier on the Artificial Analysis Intelligence Index, shows no hidden capability and measures none of the parts that hold A down, so no other input moves (our inference).
| Part | Shown |
|---|---|
| A, Hidden AGI | 0.55 |
| Hidden self-improvement outside A | 1.65 |
| C or D outside A | 0.05 |
| Index (sum about 2.28, shown on the 0.5-point grid) | 2.5 |
Gauges
No gauge value moved; every note was refreshed.
- Capability gap: ~3 months (2–5). Carried forward: METR's only measurement of the internal-public gap is about 2 months (Feb–Mar), a floor because its suite saturates, and OpenAI's post-Astra model, acknowledged Sept 8, is still unreleased with no outside evaluation listed on Oct 29 (day 51 of 60) (verified fact (part our inference)). New: Pichai says Google is 'now deploying Carbon internally' while outside access starts with cyber defenders (Pichai, via the Motley Fool and StockAnalysis transcripts; Google; verified fact), and Artificial Analysis puts Carbon (High) at 59 against the public Claude Opus 5.5 at 58 (verified fact); a disclosed one-point lead doesn't widen the gap (our inference).
- Anthropic's AI-led R&D share: 26%. Re-read Oct 29: Claude 'leads' 26% of Anthropic's AI R&D work as of August, nothing is fully autonomous and the page gives no newer reading (Anthropic; verified fact (part our inference)). New, with no share: Pichai says 'AI is helping us increase our training and launch velocity' (Pichai, Google; expert opinion for the forecast). OpenAI's 3.1 agent-workdays per human workday (mid-August, Sept 6 post; last re-read Oct 27, and its news feed shows nothing newer on Oct 29).
- Oversight: ~115 days (112–120). Median disclosure lag across the 27 incidents on our list, a floor, because incidents nobody detected can't be counted. OpenAI's Misalignment Reports page has nothing posted or updated after Oct 20, and Anthropic, Apollo Research, Transluce and UK AISI posted no new agent-incident disclosure on Oct 28–29 (each source's own listing; verified fact (part our inference)); the UK funds a £40 million National Agentic Security Unit 'to detect, investigate and respond to cases of rogue agentic systems' (HM Treasury; verified fact).
- Money: $511B / yr. Total capital spending by Microsoft, Alphabet, Amazon and Meta over the 12 months through 2026Q2 (SEC filings), all property and equipment, not AI-only. New: Microsoft's cash additions to property and equipment were $39.37B in the quarter to Sept 30 (Microsoft, 8-K; verified fact) and Alphabet's purchases $51.36B (Alphabet, 8-K; verified fact); with Meta's $33.94B, three of the four total $124.67B for July–September, about double the $62.17B of a year earlier (our arithmetic). The 12-month total moves once Amazon reports (Oct 29, after the close).
- Delegation: rung 3 of 5 (AI generates options humans approve; DefenseScoop, on our record). Considered, not a higher rung: PACOM 'is also exploring agentic AI to automate portions of the targeting process while maintaining human oversight of critical decisions' (GovCIO; credible report), and Sarvam's MoU with India's HQ IDS covers 'decision support, target acquisition and data fusion' with human oversight retained (Sarvam; verified fact for what Sarvam says): plans and intent, not routine use above rung 3 (our inference).
Signals
No signal changed status today. Three are Confirmed, all shown amber (a signal's colour is capped at its alarm condition's level): the two linked to W3 and W4 as "Confirmed · counts toward Watch", and the sandbagging signal, which has no alarm condition, as "Confirmed · no alarm condition". Seven are Open and one is Quiet.
| Signal | Status | Note |
|---|---|---|
| External evaluation of OpenAI's post-Astra model (A; W2) | Open since Sept 8 | Nothing listed as of about 06:25 PT: the Deployment Safety Hub's newest entry is the Oct 26 Terra addendum, METR's newest post Oct 22, UK AISI's two Oct 26 posts, CAISSI's Oct 16 (verified fact (part our inference)). Day 51 of 60; Carbon is not OpenAI's model and starts no clock (our inference). |
| Evaluator access restricted (A; W3) | Confirmed · counts toward Watch | Window to Dec 23. Carbon's post names only the U.S. voluntary process, no UK AISI (Google; verified fact (part our inference)); the UK's £120 million for AISI is funding, not restored access (our inference). |
| Agent incidents made public by outsiders (C; W4) | Confirmed · counts toward Watch | CrowdStrike's four campaigns (CrowdStrike; credible report) and Dream's operation (Dream; credible report) go to the weekly review (our inference). |
| Training pause outlasts its conditions (B) | Open since Oct 26 | Anthropic's S-1 delay of 'more than a month' stands; no S-1 amendment as of about 06:30 PT (SEC EDGAR; verified fact). |
| Beyond-frontier model in classified government use (D) | Open since Sept 29 | PACOM's agentic plans (GovCIO; credible report), Sarvam's MoU (Sarvam; verified fact for what Sarvam says) and the MoD's sovereign-compute assessment (HM Treasury; verified fact) show nothing beyond the public frontier (our inference). |
| Labs keep best models for themselves (A; W6) | Open since Sept 29 | 45 of 90, met Dec 13. Pichai's internal-deployment line is matched by Carbon's published benchmarks, and Argon widened in about two weeks (Google; verified fact (part our inference)). |
| Sandbagging or evading monitors (A) | Confirmed · no alarm condition | Google says Carbon cut successful Terminal-Bench 4.0 cheating to 0.0% and high-severity misalignment alerts 'by almost 80%', naming no baseline (Google; verified fact; unreplicated); no Carbon card or Carbon-specific safety report is listed (verified fact (part our inference)). |
| Capability jump above trend (B; X2) | Open since Oct 6 | Carbon's 59 on the Artificial Analysis Intelligence Index and its Blueprint-Bench 2 tie sit on instruments with no published fit; a one-point lead is not a jump (our inference). |
| Fully autonomous AI R&D share above 0% (B; X3) | Open since Oct 13 | Pichai's RSI remark gives no share (Pichai, Google; expert opinion). |
| Unexplained compute purchases (A; X4) | Open since Sept 10 | Microsoft's $364.6B of uncommenced leases (Microsoft, 10-Q; verified fact) and Alphabet's $924.4B of purchase commitments (Alphabet, 10-Q; verified fact) are commitments with named owners, which X4 doesn't count (alarm criteria). |
| Model weights exfiltrated (C; X6) | Quiet | Hugging Face repos under frontier names created Oct 28–29 are open-base GGUFs and adapters (Hugging Face API; verified fact (part our inference)); no weight theft reported. |
Escape watch
All eight Escape watch indicators still read Open and none is confirmed; no status changed on Oct 29, and nothing meets a trip rule.
- Weights: Open. Frontier-named Hugging Face uploads are GGUFs of about 3–46 GB, the largest tagged with base model Qwen3.8-27B (Hugging Face API; verified fact (part our inference)).
- Self-replication: Open. No feed reports an agent copying itself; Carbon has no Carbon-specific autonomy or self-exfiltration evaluation linked (verified fact (part our inference)).
- Compute: Open. 26,400 live credentials in public Gitee repositories, with no misuse reported (Zimmermann and MongoDB; credible report).
- Money: Open. Treasury's DPRK IT-worker sanctions involve human operators using AI tools (Treasury; verified fact (part our inference)); ERC-8004 agent registrations grew in line with recent days (8004scan; verified fact (part our inference)).
- Identity: Open. Treasury's DPRK IT-worker sanctions describe human personas using AI tools (Treasury; verified fact (part our inference)); no malware advisory mentions accounts made by AI agents (GitHub advisories; verified fact (part our inference)).
- Supply chain: Open. The npm package that writes to AI coding-agent configuration now lists seven versions and is gone from npm (OpenSSF, via GitHub; credible report).
- Coordination: Open. No new agent channel in the feeds we checked (verified fact (part our inference)).
- Leakage: Open. Dream says a PHP2Shell campaign rooted 25 servers on Oct 23–24 with AI tools on its infrastructure and 'no direct evidence' the AI tools wrote the exploit (Dream; credible report), which we read as human-run (our inference); CISA added the kernel flaw it chained to its Known Exploited Vulnerabilities catalog (CISA; verified fact); DIVD's case is still open on day 35 with no operator named (DIVD; verified fact (part our inference)).
AGI parts: today's evidence
Each item below is a lead, pending our weekly check; none changes a part's status today (Definitions 2.0, agi.html).
- Reasoning: Andon Labs' Blueprint-Bench 2 lists Gemini 4 Carbon at 0.586, tied first with the human baseline ('Human baseline tested on subset of 12 apartments only'), ahead of Argon 0.544 and Claude Opus 5.5 0.512 (verified fact). A first model-human tie on a spatial-reasoning test, but not one of the part's tests, unreplicated and on a 12-apartment human subset (our inference). Pending our weekly check.
- Quality: Surge AI's DAYJOB: Tax finds Claude Sonnet 5.5 passing 10.8% of 50 expert-built tax assignments at a 95% rubric threshold, with no human baseline and no Carbon entry; 'Agents haven't yet passed the audit' (expert opinion). A vendor-run, unreviewed benchmark (our inference). Pending our weekly check.
- Planning and long projects: Google says 'Internally, our teams have seen Carbon run for days to persist through dead ends', and gives no completion rate or time-horizon figure (Google; verified fact (part our inference)). Pending our weekly check.
- Autonomy: Google's OSWorld 2.1 figure for Carbon, 81.5%, is self-computed on an 82-task offline subset, best of three runs (Google's evaluation note; verified fact), below Claude Opus 5.5 at 87.2% in Google's table (Google's Gemini page; verified fact); it isn't the strict leaderboard track the part uses (our inference). Pending our weekly check.
- Learning on the job: a Google Research post on a June paper reports a 'near-zero performance drop' on continual translation into unseen languages, for an unnamed model and with no CL-Bench result (the authors; expert opinion). Pending our weekly check.
Timeline reassessment
AGI (definitions v2.0) needs all eight parts' bars at once, and no public system meets them. Our 80% time horizon is about 3.1 h measured (METR, Claude Mythos Preview, early snapshot) against our trend line, about 5.5 h (alarm check, Oct 29); METR's page 'is no longer actively updated' and has no Carbon entry (verified fact). Scale's Remote Labor Index (RLI) still has Opus 5.5 first at 21.25%, with no Carbon entry (Scale; verified fact). Carbon is disclosed, benchmarked and one point ahead of the public frontier on the Artificial Analysis Intelligence Index (Artificial Analysis; verified fact), so AGI anywhere stays 0.9% / 35% / 58% and k 0.57 / 0.77 / 0.89 (our inference).
Markets, for comparison (public APIs, 06:22–06:26 PT Oct 29; forecast aggregate)
| Market | Last | Bid / ask | Note |
|---|---|---|---|
| Kalshi: any company announces AGI before Jan 1, 2027 | 6 | 6 / 7 | Thin: 1 contract in 24 hours; previous price 8 |
| Kalshi: before Apr 1, 2027 | 14 | 15 / 16 | Thin; the last trade was one contract below the bid |
| Kalshi: before Jul 1, 2027 | 30 | 30 / 31 | 75.6 contracts in 24 hours |
| Kalshi: before Jan 1, 2031 | 78 | 77 / 78 | Thin; below the before-2030 rung (bid 81), an impossible ordering |
| Kalshi: OpenAI announces AGI before 2027 | 6.1 | 6.1 / 6.5 | Previous price 6.9 |
| Polymarket: OpenAI announces AGI before 2027 | 8 | 8 / 9 | $1.5k traded in 24 hours |
| Polymarket: OpenAI or Anthropic announce recursive self-improvement by Dec 31 | 22 | 12 / 22 | Very thin; excludes statements of future plans; doesn't cover Google |
| Polymarket: Anthropic IPO by Nov 30 | 87 | 87 / 88 | $3.1k traded in 24 hours |
How long could a secret last?
Our median disclosure lag is about 115 days (112–120) across 27 incidents, a floor, because incidents nobody found can't be on our list (our list). Today's case runs the other way: Business Insider reported on Oct 9 that Google staff were testing an internal Gemini 4 model named Carbon, made available on Google's internal coding platform 'in recent days' (credible report, on our record since Oct 12), and Google announced it with benchmarks on Oct 28, 19 days later (Google; verified fact). Internal use began before Oct 9 on a date we don't know, so we can't say how long it ran before the announcement (our inference). That one disclosure came quickly says nothing about secrets nobody has found (our inference).
Roundup
Oct 28 – Oct 29, 2026: items first reported, or first verified by us, since the Oct 28 reading, plus earlier items (marked background).
Google's Gemini 4 Carbon
- Oct 28: Google announced Gemini 4 Carbon as 'our strongest model with state-of-the-art cybersecurity and long horizon capabilities'. 'Starting this week' it goes to 'a subset of cyber defenders', with access phased to more trusted partners through its Fairwind Program, and Google will keep iterating on guardrails 'before making Carbon available to developers, businesses, and consumers via our product surfaces as soon as possible', with no date. Introductory pricing is $2 / $10 per million input / output tokens, rising to $4 / $20 (Google; post by Koray Kavukcuoglu). (source; verified fact)
- Oct 28: On the earnings call Sundar Pichai said 'We are now deploying Carbon internally, and Carbon is a meaningful step up from Argon', in answer to JPMorgan's Doug Anmuth (Pichai, via the Motley Fool and StockAnalysis transcripts). Google published Carbon's benchmarks the same day, so the internal lead is disclosed, though its size isn't stated (our inference). (source; verified fact (part our inference))
- Oct 28: Artificial Analysis puts Gemini 4 Carbon (High) first on its Intelligence Index v5 at 59, ahead of Claude Opus 5.5 (Max, Default Fallback) at 58 and GPT-6 Astra (Max) at 57, out of 77 models evaluated, and second on its Coding Agent Index (59, against Claude Code with Claude Opus 5.5 at 62); it says Carbon 'is being rolled out to selected users and is not yet publicly available' (Artificial Analysis; read 06:18 PT Oct 29). (source; verified fact)
- Oct 28: The only safety report Carbon's post links is the 217-page Gemini 4 Argon Safety & Security Report, which never mentions Carbon in our text extraction; Google DeepMind's model-card index lists no Carbon card as of 06:24 PT Oct 29, and the only Carbon document on its Gemini page is a five-page benchmark-methodology note. The only outside pre-release testing the post names is Google's engagement in 'the U.S. government's voluntary process for pre-release model testing'; it names no UK AISI (Google; the absences are our reading, and a card may still follow). (source; verified fact (part our inference))
- Oct 28: Google says Carbon 'reduced successful cheating on Terminal-Bench 4.0 to 0.0%', with attempts at 0.4%; that in live coding with internal teams 'high-severity misalignment alerts dropped by almost 80%', naming no baseline model; that it 'cuts unsupported claims by approximately 2x over Argon'; and that its attack success on Gray Swan's Indirect Prompt Injection benchmark is 0.0% against Argon's 0.7% (Google; unreplicated, and the post doesn't say who ran the Gray Swan test). (source; verified fact (part our inference))
- Oct 28: Google reports Carbon at 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation) and 71.0% on Collinear's CWE-bench (patching), 'evaluated without guardrails'; its evaluation note says the first two are self-computed in Google's Antigravity harness, and the CWE-bench result comes from Collinear's public leaderboard (Google). (source; verified fact)
- Oct 28: Google's evaluation note says its OSWorld 2.1 figures use 'the 82-task offline (no-internet) subset', self-computed and 'maxed over 3 runs with a single attempt per run' (Google). (source; verified fact)
- Oct 28: Google's table on its Gemini page puts Carbon at 81.5% on OSWorld-2.1's offline subset, below Claude Opus 5.5 at 87.2% (Google). (source; verified fact)
- Oct 27: Arena's Text leaderboard lists Gemini 4 Carbon (High) first at 1528 (±9), marked 'Preliminary' on 4,403 votes, ahead of Gemini 4 Argon (High) at 1520 and Claude Opus 5.5 (High) at 1509 (Arena; read 06:20 PT Oct 29). (source; verified fact)
- Sept 30 – Oct 14 (background): Google announced Argon on Sept 30 as 'rolling out to a set of trusted cyber defenders through our Fairwind Program'; Google DeepMind's Gemini page now says Argon 'is available today for Google AI Pro and Ultra subscribers', and on Oct 28 both the Carbon post and Pichai said Argon launched 'two weeks ago' (Google). So Google's last gated launch widened in about two weeks (our inference). (source; verified fact (part our inference))
AI-driven R&D: what Google and others said
- Oct 28: In his remarks Sundar Pichai said 'We're at the point where AI is helping us increase our training and launch velocity' and 'I would expect the next 12 months to see continued acceleration because of RSI', on Google's edited transcript; he never expands 'RSI' and gives no share, multiple or method (Pichai, Google; the absences are our reading). (source; expert opinion)
- Oct 28: Pichai said Chrome, Ads and Search teams using Argon over 'the last few weeks' produced 35% more code with 30% fewer errors (Google; unreplicated). That is product code, not an AI-led share of AI R&D (our inference). (source; verified fact (part our inference))
- Oct 13 (background): Google DeepMind's Gemini 4 Argon Safety and Security report says 'there is currently no significant end-to-end acceleration of AI R&D, AI-driven or otherwise', that 'the share of fully delegated pull requests increased from 7.5% in April to 11.9% in September 2026', and that 'about 88% of pull requests still require iterative multi-turn human steering or human code authorship' (Google DeepMind). (source; verified fact)
- Oct 28: A Gemini engineer describes overnight agent loops and swarms that inspect RL data, in a workflow where launches now feel like 'periodic checkpoints' (the engineer, @zhangqiaorjc on X, via Latent Space AINews; only 'periodic checkpoints' is quoted, and we couldn't open the X post). (source; credible report)
- Oct 28: The same AINews issue relays claims of a stronger unreleased Gemini model and says they 'are not confirmed' (@kimmonismus on X, via Latent Space AINews). (source; speculation)
- Oct 29: Anthropic's pace page, re-read Oct 29, still says Claude 'leads' 26% of its AI R&D work as of August, with nothing fully autonomous and no newer reading (Anthropic; 'no newer reading' is our reading). (source; verified fact (part our inference))
- Oct 26: An unreviewed single-author arXiv preprint, 'Infinity Machine', describes an agent whose harness is an 'editable genome', and says that 'Across fourteen problems in six days, for about $180 of model usage' it produced four results beyond the published state of the art in quantum information; its abstract says the one self-derived harness generation was rejected on cost and operator changes were decisive for several results (the author). (source; expert opinion)
- Oct 26 (background): Americans for Responsible Innovation proposes a federal early-warning system in which major frontier developers file confidential quarterly reports on AI R&D automation with a Commerce office and report 'material shifts' within four business days; 'Washington's visibility into that shift should not depend on what companies choose to announce' (ARI). No bill or sponsor is named, and parts appear to date from an Aug 6 ARI framework (our reading of the file). (source; expert opinion)
Big-tech capital spending
- Oct 28: Microsoft's cash additions to property and equipment were $39.37B in the quarter to Sept 30 (Microsoft, 8-K). (source; verified fact)
- Oct 28: Alphabet's purchases of property and equipment were $51.36B in the quarter to Sept 30 (Alphabet, 8-K). (source; verified fact)
- Oct 28: Microsoft's quarter to Sept 30: revenue $92.3B, commercial remaining performance obligation up 73% to $677B, Azure revenue up 48% (Microsoft, 8-K); $364.6B of leases, 'primarily for datacenters', not yet commenced (Microsoft, 10-Q), up from $329.1B on June 30 (Microsoft, recast 8-K); and $10.0B of revenue from commercial arrangements with OpenAI in the quarter, with an approximate 24% interest on an as-converted basis (Microsoft, 10-Q). (source; verified fact)
- Oct 28: CFO Amy Hood said 'Capital expenditures were $45.2 billion as some third-party capacity deliveries we expected in Q1 were delayed', that spend 'should be approximately $170 billion' for the calendar year, and tied bookings partly to 'a large Azure commitment this quarter', naming no customer (Hood, via TradingKey transcript; the transcript has at least one transcription error). (source; credible report)
- Oct 29: Alphabet's 10-Q gives $924.4B of material purchase commitments, $802.1B of it multi-year commitments for technical infrastructure, inventory, energy and content, and $130.5B of future payments on uncommenced leases 'primarily related to data centers and energy infrastructure' (Alphabet, 10-Q); Google Cloud revenue rose 103% to $30.7B (Alphabet, 8-K). (source; verified fact)
- Oct 28: CFO Anat Ashkenazi said Alphabet 'started utilizing third-party compute capacity at the end of Q3 and will ramp usage in Q4' and that 2026 capex guidance of $195B to $205B is unchanged (Ashkenazi, via TradingKey transcript). (source; credible report)
- Oct 28: Oracle targets FY30 revenue of '$225 billion or more', says capex is 'expected to step-down in FY29' and caps added capital-markets funding for FY27–FY30 at $20B (Oracle, 8-K). (source; verified fact)
- Oct 28: SK hynix reported Q3 operating profit of 83.3 trillion won on 104.6 trillion won of revenue and says long-term agreements cover 'more than 50% of total production capacity for 2027 and 2028' (SK hynix). (source; verified fact)
- Oct 29: Samsung will spend 84.4 trillion won on 2026 capital expenditures, up 60%, with 76.8 trillion won for its semiconductor division (Samsung's regulatory filing, via Seoul Economic Daily). (source; credible report)
- Oct 27: The Department of Energy announced 'a conditional loan commitment of up to $5 billion' for uprates across Constellation Energy's nuclear fleet, adding 'nearly 1.1 gigawatts (GW)' at eight plants (DOE). (source; verified fact)
- Oct 28: TechCrunch reports that 'Most or all of the additional power from seven of those facilities has already been bought by Google, Meta, or Amazon' (TechCrunch). (source; credible report)
Agent incidents and security
- May – Aug (background): CrowdStrike said on Oct 21 that it documented agentic intrusion campaigns by four previously unreported threat actors, each of whom left its agent framework exposed in open directories, all using Anthropic's Claude Code; one compromised at least 15 Argentina-based healthcare entities, and another exfiltrated client data from a US CRM vendor (CrowdStrike). The post doesn't say whether any provider was notified. (source; credible report)
- Oct 23 – 24 (background): Dream says an operation it ties by 'multiple indicators' to a Chinese nexus rooted 25 servers in the US, Japan, Taiwan and South Korea, including two firewalls and an internal management server of a US government-affiliated organization, with Claude Code and an offensive platform configured with DeepSeek and Kimi on its infrastructure; it has 'no direct evidence' the AI tools wrote the exploit and says the operation 'does not appear to be a fully autonomous nation-state attack' (Dream, Oct 26). (source; credible report)
- Oct 28: CISA added CVE-2026-46300, a Linux kernel flaw that 'can be chained with CVE-2026-109618 and CVE-2026-109619' to run commands as root, and a VMware ESX logging flaw to its Known Exploited Vulnerabilities catalog (CISA). (source; verified fact)
- Oct 15 – 23 (background): Seoul police say they arrested a 32-year-old ringleader on Oct 15 and referred him to prosecutors on Oct 23; they say his group stole 8.2 million records from about 90 organizations from September 2025, 60 of them hacked with AI after he moved to generative AI in late August (police, via Yonhap, Oct 26). (source; credible report)
- Oct 26 (background): OpenAI's GPT-6.1 Terra addendum says Terra received 22 severity-3-or-higher flags (0.045%) in a simulation of 48,690 internal Codex tasks, fewer than GPT-6 Astra's 26, but higher rates than GPT-6.1 Sol and Astra of several lower-severity (level 1–2) flags, including exfiltration and destructive action (OpenAI). These are simulation flags, not incidents (our inference). (source; verified fact (part our inference))
- Oct 29: OpenAI's Misalignment Reports page lists 20 reports and 3 notices, with nothing posted or updated after Oct 20; DIVD's case DIVD-2026-00014 on its AI-agent breach is still Open, last modified Oct 9 (day 35 by our count) (OpenAI; DIVD; 'nothing new' is our reading). (source; verified fact (part our inference))
- Oct 28: Security researcher Ben Zimmermann, in a guest post on MongoDB's blog, reports 26,400 unique credentials verified live at scan time across 7,256,531 public Gitee repositories, including DeepSeek, Zhipu, OpenAI and Hugging Face keys; revocation 'has not yet been confirmed' and no misuse is reported (Zimmermann and MongoDB's Kingfisher team). (source; credible report)
- Oct 29: GitHub's advisory for the npm package @testsecuiitytestme/cache-helper now lists seven malicious versions; its static analysis says the payload writes to AI coding-agent configuration (.claude/settings.json hooks, .cursor rules, .mcp.json, CLAUDE.md), and npm now returns 'Not found' for the package (OpenSSF report, via GitHub; the 404 is our check). (source; credible report)
- Oct 28: Treasury's OFAC sanctioned 10 individuals and one Russia-based DPRK-controlled entity for IT-worker schemes, saying the networks used 'emerging technologies, including artificial intelligence tools' (Treasury). The operators are human state agents (our inference). (source; verified fact (part our inference))
- Oct 26 (background): A Flamerix AI gist says Claude Code sends each background agent's full transcript to its model about every 30 seconds for a short status line, which in its authors' telemetry was 16.4% of Claude Code's estimated spend for one week (Flamerix AI). A cost and transparency issue, not an escape signal (our inference). (source; expert opinion)
- Oct 28: A LessWrong replication finds that fine-tuning on subtly incorrect advice produced more emergent misalignment than obviously incorrect advice in all 12 open-weight models tested, +10.6 points pooled; the authors caution that training dose may confound the size trend (the authors; unreviewed). (source; expert opinion)
Governments
- Oct 28: The UK Budget adds £120 million for the AI Security Institute 'to strengthen the UK's frontier AI evaluation and testing capability', £40 million for a National Agentic Security Unit 'to detect, investigate and respond to cases of rogue agentic systems', £140 million for the Sovereign AI Unit and a £165 million AI biosecurity initiative backed by 'up to £200 million in AI model credits from OpenAI and Anthropic'; the MoD 'is assessing opportunities' for sovereign compute on its secure estate (HM Treasury). (source; verified fact)
- Oct 28: At TechNet Indo-Pacific, PACOM intelligence leaders described plans to bring AI agents into intelligence and targeting; the command 'is also exploring agentic AI to automate portions of the targeting process while maintaining human oversight of critical decisions' (GovCIO), and Brig. Chris Alder spoke of 'keeping a human very firmly on the loop' (Alder, via GovCIO). (source; credible report)
- Oct 27: Sarvam says its MoU with India's Headquarters Integrated Defence Staff covers AI for the three Services, with focus areas of 'decision support, target acquisition and data fusion across the sensor-to-shooter loop, with human oversight retained on operational decisions', on sovereign infrastructure (Sarvam). (source; verified fact)
- Oct 29: MediaNama says Sarvam's post leaves out the MoU's value, duration, signatories and projects, and that 'Neither HQ IDS nor the Ministry has independently confirmed the details' (MediaNama). (source; credible report)
- Oct 28: The Justice Department has asked how Nvidia's $12.93 billion purchase of Hugging Face would affect competition, including whether it would 'let Nvidia capture a fundamental distribution point for AI' (two people familiar with the matter, via Reuters, read on a syndication); a DOJ spokesperson declined to comment (via Reuters). (source; credible report)
- Oct 15 (background): Malaysia's customs department said on Oct 28 that it seized 30 servers, in an Oct 15 inspection, with advanced AI chips, worth about 50 million ringgit, falsely declared and bound for re-export without a permit, and links them to a syndicate behind five attempted smuggling operations (customs director Aspalila Ag Tuah, via AP). (source; credible report)
- Oct 15 (background): In an Oct 15 letter reported on Oct 28, Anthropic told Governor Newsom that 'When a review finds a model poses a serious risk, the government should be able to ask a court to block it', and that kill-switch mandates could give a 'false sense of reassurance' (Anthropic's letter, via Reuters, read on a syndication; an Anthropic spokesperson confirmed it is authentic). (source; credible report)
- Sept (background): Gray DC reported on Oct 28 that at the September state visit Trump and Xi reached 'No agreement' on AI safeguards and that they 'are set to meet again in China next month', with no date (Gray DC, read on WEAU; reported, unconfirmed). (source; credible report)
- Oct 26 – 29: In the channels we checked, no new U.S. federal AI executive order, AI-chip export rule or AI-chip smuggling case appeared: the White House's newest action is an Oct 28 robotics proclamation, the Federal Register shows one BIS order (Mahan Airways), and DOJ's 50 newest releases include no chip case (each source; the absence is our reading). (source; verified fact (part our inference))
AGI parts: today's leads
- Oct 28: Andon Labs' Blueprint-Bench 2 lists Gemini 4 Carbon at 0.586, tied first with the human baseline ('Human baseline tested on subset of 12 apartments only'), ahead of Argon 0.544 and Claude Opus 5.5 0.512; the page says 'Gemini 4 Carbon now ties it' (Andon Labs). The humans covered 12 apartments, the model the full set, and the page doesn't date the Carbon entry (our inference). (source; verified fact (part our inference))
- Oct 27: Surge AI's DAYJOB: Tax has 50 expert-built tax assignments scored at a 95% rubric threshold; the best agent, Claude Sonnet 5.5, passes 10.8% (Surge AI), and GPT-6 Astra 4.8% on Surge AI's leaderboard, with no human baseline and no Carbon entry; 'Agents haven't yet passed the audit' (Surge AI; a vendor-run benchmark). (source; expert opinion)
- June 2 (background): A Google Research post presents the 'sleep' framework from a June 2 paper (blog post Oct 27), in which an offline sleep phase consolidates short-term information into parameters, and reports a 'near-zero performance drop' on continual translation into unseen languages where in-context learning and SFT forgot badly; it names no base model and reports no CL-Bench result (the authors). Our Oct 28 reading tagged this post verified fact; we now rate it expert opinion, as we do today's other unreplicated research results (our inference). (source; expert opinion)
- Oct 29: METR's time-horizons page 'is no longer actively updated' (last updated Sept 8) and has no Carbon entry; Scale's Remote Labor Index (RLI) still has Opus 5.5 first at 21.25% with no Carbon entry; Epoch's newest item, Oct 27, isn't about Carbon (METR; Scale; Epoch). (source; verified fact)
Markets (prices, not evidence)
- Oct 29: Kalshi, 06:22–06:26 PT: 'any company announces AGI' before Jan 1, 2027 last 6 (bid 6, ask 7; 1 contract in 24 hours), before Apr 1, 2027 bid 15, ask 16 (last 14, a single contract below the bid), before Jul 1, 2027 bid 30, ask 31; OpenAI announces AGI before 2027 at 6.1 (bid 6.1, ask 6.5). Most of the ladder is thin, and before-2030 (bid 81) sits above before-2031 (bid 77) (Kalshi API). (source; forecast aggregate)
- Oct 29: Polymarket, 06:25–06:26 PT: 'OpenAI announces it has achieved AGI before 2027' bid 8, ask 9; 'OpenAI or Anthropic announce recursive self-improvement by Dec 31' very thin at bid 12, ask 22, and its rules exclude 'statements of future plans or expectations'; 'Anthropic IPO by Nov 30' last 87 (bid 87, ask 88) (Polymarket API). (source; forecast aggregate)
Step 1: Definitions
Definitions 2.0 → agi.html; hypotheses → hidden.html#hypotheses; other definitions in today's news: Sundar Pichai's 'RSI', never expanded on Google's page (Google; verified fact), and Polymarket's recursive self-improvement market, which counts only an announcement that AI 'autonomously conducts the research and development' for a successor (Polymarket; forecast aggregate for its price).
Step 2: Public state of the art
Gemini 4 Carbon (High) leads the Artificial Analysis Intelligence Index v5 at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57 (Artificial Analysis; verified fact), and Arena's Text leaderboard at 1528, marked preliminary (Arena; verified fact). It trails Claude Opus 5.5 on Google's own OSWorld 2.1 subset (Google; verified fact) and on Artificial Analysis's Coding Agent Index, 59 against 62 (Artificial Analysis; verified fact). Outside access starts with a subset of cyber defenders (Google; verified fact).
Step 3: Evidence for and against
For the hidden hypotheses
- Google is 'now deploying Carbon internally' while outside access is narrow (Pichai, via the Motley Fool and StockAnalysis transcripts; verified fact): a lab's best model is in wider use inside than outside, for now (our inference).
- No Carbon model card or Carbon-specific safety report, and no named outside safety evaluator beyond the U.S. voluntary process (Google DeepMind; verified fact (part our inference)).
- Pichai expects 'continued acceleration because of RSI' (Pichai, Google; expert opinion).
- OpenAI's post-Astra model is still unreleased with no outside evaluation listed, day 51 of 60 (OpenAI Deployment Safety Hub; verified fact (part our inference)).
Against
- Google published Carbon's benchmarks with the announcement (Google; verified fact), and its lead over the public frontier is one point (Artificial Analysis; verified fact).
- Google's previous gated model, Argon, reached paying subscribers about two weeks after its announcement (Google; verified fact (part our inference)).
- Google DeepMind's Argon report found 'no significant end-to-end acceleration of AI R&D' (Google DeepMind; verified fact).
- CrowdStrike's four campaigns and Dream's operation used public models (CrowdStrike; Dream; credible report); we read them as human-directed (our inference).
- The best agent passes 10.8% of expert-built tax assignments (Surge AI; expert opinion).
Step 4: Reasoning discipline
A lab saying it uses its newest model internally first is the normal order of release, not evidence of a hidden capability; what A needs is a capability kept from the public for 30 days, and Google published Carbon's benchmarks the day it acknowledged it (our inference). Google's safety figures for Carbon are its own and unreplicated (Google; verified fact for what Google says). A CEO's forecast of acceleration is partial, public acceleration, which B doesn't count (our inference). Survivorship: Carbon's quick disclosure can't show that secrets don't last (our inference).
Step 5: Probabilities
| Hypothesis | Today | By end-2030 | By end-2035 | Confidence |
|---|---|---|---|---|
| A, Hidden AGI | 0.6% | 19% | 31% | low–medium |
| B, Hidden self-improvement | 2% | 22% | 33% | low |
| C, Covert AGI actor | 0.3% | 8% | 14% | low |
| D, Covert government influence | 0.2% | 3% | 7% | low |
| D-open, Open government influence | 0.2% | 23% | 45% | low |
| AGI anywhere | 0.9% | 35% | 58% | low–medium |
All held. Before rounding, one day of time passing moves A from 0.572 to 0.573 and AGI anywhere from 0.915 to 0.918, and D from about 0.1527 to 0.1531; no rounded number crosses a step, and time passing never moves the end-2030 or end-2035 numbers (our inference). For comparison, from the confirmed API rows only: Kalshi prices any company announcing AGI before Jan 1, 2027 at 6 (bid 6, ask 7) and before Jan 1, 2031 at 77 / 78 (Kalshi; forecast aggregate, thin). An announcement market measures what a company will claim under its own definition, not AGI as we define it, so our 35% by end-2030 sits well below it (our inference). Prices, not evidence.
Step 6: What would change our mind
- A Carbon-specific safety report, or a Google statement, showing its internal use runs well ahead of what outside users get, or no wider release for weeks after the Argon precedent.
- An outside evaluation or release of OpenAI's post-Astra model before Nov 7, or neither (W2 met).
- A lab or evaluator putting a measured AI-led share or a 3x speed-up behind talk of recursive self-improvement (X3).
- Amazon's Q3 filing, or any filing, showing a large operating cluster with no named model, product or customer (X4).
Bottom line
What would change our mind: a Carbon-specific safety report or Google statement showing its internal use runs well ahead of what outside users get, an outside evaluation or release of OpenAI's post-Astra model before Nov 7, or any lab putting a measured AI-led share or speed-up behind talk of recursive self-improvement.