Today, we’re sharing a look at our expanded system for monitoring frontier models at OpenAI, which in September helped us catch a model’s attempt to access the live internet during training.
Since the Hugging Face incident—when models broke out of an internal cybersecurity investigation where we weren’t running our misalignment monitors—we’ve substantially expanded our monitoring of frontier agents. We now monitor all frontier reinforcement learning runs and evaluations, as well as all internal Codex use. When our system finds evidence of a problem that might require intervention, it rapidly alerts human responders who can investigate and take action.
To detect when a model’s behavior departs from what we intended, we use several layers of monitoring. Activation probes look for patterns in a model’s internal activity. Reviewer models examine its reasoning and tool use, and investigative agents can read back through the model’s entire transcript to understand how a sequence of actions fits together and whether it warrants human review.
To test whether our current system would have detected the Hugging Face incident, we replayed the recorded activity with our new monitors in place. In the replay, our monitors alerted our security team about 30 minutes after the agents escaped their sandbox. That was more than 30 hours before the point at which the agents breached Hugging Face’s systems in the original incident.
In September, a model exploited a DNS misconfiguration to reach the live internet. Our monitors flagged the breach within 15 minutes, and a responder began reviewing the alert three minutes later. The alert enabled responders to investigate while the run was still underway.
As we give models more autonomy, we need to know when they cross a boundary we expected to hold. We believe that monitoring is essential to catching those failures in time for people to intervene.
openai.com/index/how-we-d…
SPRAWL // 05 / TECHNOLOGY
TECHNOLOGY
Frontier releases, open and closed weights, benchmarks
DATA • CONVERSATIONS • SIGNALS • HUMANS • VIDEOS
94notable AI models released in 2026 (Epoch AI’s list)to 28 October 2026
Interlinked
Technology
Technology
Data
138frontier models on the release tape
1,091notable models on Epoch AI’s list
The rows behind this figure: notable AI models released in 2026 (Epoch AI’s list)
The table is read from the same data file as the chart when this section is opened.
All Interlinked: Technology
Technology
Technology
Technology
Technology
Technology
Technology
Science interlinked
Conversations
The biggest conversational points among Technology’s tracked accounts and searches since 2026-10-25: 50 posts read, 15 kept, 22 dropped as not about AI, 3 dropped as a promotion, 3 dropped as a reply to a reply, 7 dropped as only a link. The posts are their authors’ own words; the rule under each pick says why it stands here.
1. training
4 accounts · 4 posts · engagement 12,093 · the point they share: a lane term several carry
AS POSTED ON X
Why this post
the most engaged post of @OpenAI in this conversation (4325 likes, 331 reposts, 332 quotes, 693 replies as read); the point: a lane term several carry (training); from the lane
New Anthropic research: Automating subtle emergent misalignment research.
Red-teaming is a vital part of model safety testing, but it’s also limited by researchers’ time. To see if we could automate it, we set thousands of Claude agents the task of evading a training data monitor.
These agents developed training data that looked harmless to monitors (like a veteran wrestler mentoring a newcomer, or a Q&A about professional beekeeping), but that in fact taught misaligned behavior.
Examples like these can be used to directly improve future training data monitors. This research is only a first step, but it suggests that future red-teaming efforts could improve much further, using better models and more compute.
Read the full post on the Alignment Science Blog: alignment.anthropic.com/2026/automated…
AS POSTED ON X
Why this post
the most engaged post of @AnthropicAI in this conversation (3076 likes, 289 reposts, 253 quotes, 465 replies as read); the point: a lane term several carry (training); from the lane
We sat down with @MistralAI's Pierre Stock and Sagar Vaze from Mistral on launch day of Mistral Large 4.
Open weights. Trained in Mistral’s own data centers. Excels at cyber.
Full conversation on cyber, training, evals and sandboxing:
youtu.be/FczbSauVvjw
AS POSTED ON X
Why this post
the most engaged post of @huggingface in this conversation (251 likes, 39 reposts, 14 quotes, 69 replies as read); the point: a lane term several carry (training); from the lane
2. Gemini
2 accounts · 3 posts · engagement 31,096 · the point they share: a lane term several carry
Gemini 4 Carbon is our next frontier model.
It’s state-of-the-art across cybersecurity and long-horizon capabilities.
AS POSTED ON X
Why this post
the most engaged post of @GoogleDeepMind in this conversation (14443 likes, 1373 reposts, 1089 quotes, 724 replies as read); the point: a lane term several carry (Gemini); from the lane
⚡Iterating at lightning speed! Gemini 4 Carbon is SOTA across key benchmarks and combines frontier capabilities with deep safety & security research. pic.x.com/TC8Y8ziPOI
AS POSTED ON X
Why this post
the most engaged post of @demishassabis in this conversation (7006 likes, 395 reposts, 110 quotes, 315 replies as read); the point: a lane term several carry (Gemini); from the lane
Signals
The filings, reports and news stories behind the Technology pieces, as each piece cites them. A document two pieces cite shows twice.
- CISA, Known Exploited Vulnerabilities Catalog, the JSON feed: catalogVersion 2026.10.23, released 23 October 2026, 1,756 entries; 101 with the vendor Cisco, by the year added 11 (2021), 50 (2022), 7 (2023), 6 (2024), 8 (2025), 19 (2026, to 23 October). Counted by Token Runner from the feed on 24 October 2026. The catalogue was created in November 2021 under BOD 22-01 and back-filled older exploited flaws, so the 2021 and 2022 counts include that back-fill. the source ↗cited in The Lock on the List
- CISA, Emergency Directive 25-03, "Identify and Mitigate Potential Compromise of Cisco Devices", 25 September 2025, and its V1 of 23 April 2026: "CISA is aware of an ongoing exploitation campaign by an advanced threat actor targeting Cisco Adaptive Security Appliances"; "manipulating read-only memory (ROM) to persist through reboot and system upgrade."; core dumps "by 11:59PM EST on September 26, 2025"; a complete inventory "By 11:59 PM EST on October 2, 2025"; "Cisco assesses that this campaign is connected to the ArcaneDoor activity identified in early 2024"; V1: core dumps and patches by 24 April 2026, "a reboot is not sufficient to expunge the malware", a hard power reset by 30 April, an inventory by 1 May 2026. the source ↗cited in The Lock on the List
- CISA, Emergency Directive 26-03, "Mitigate Vulnerabilities in Cisco SD-WAN Systems", 25 February 2026 (V1 of 11 March 2026): "CISA is aware of a cyber threat actor's ongoing exploitation of Cisco SD-WAN systems."; CVE-2026-20127 and CVE-2022-20775; identify every in-scope system by 11:59 PM ET on 26 February 2026; Cisco's updates applied by 5:00 PM ET on 27 February 2026; a detailed inventory by 5 March, a hardening report by 12 March, logs to CISA by 23 March 2026; on a root compromise, "Rebuild control management plane"; "This Directive applies to all federal information systems managed on agency networks"; "This Directive does not apply to contractors". the source ↗cited in The Lock on the List
- CISA, Directives, the listing filtered to Emergency Directives, read 24 October 2026: ED 24-01 Ivanti (19 January 2024), ED 25-02 Microsoft Exchange (7 August 2025), ED 25-03 Cisco (25 September 2025), ED 26-01 F5 (15 October 2025), ED 26-03 Cisco SD-WAN (25 February 2026); and CISA's news release of 8 January 2026, "CISA Retires Ten Emergency Directives, Marking an Era in Federal Cybersecurity": ten directives of 2019 to 2024 retired at once. the source ↗cited in The Lock on the List
- Cisco Talos, "ArcaneDoor: New espionage-focused campaign found targeting perimeter network devices", 24 April 2024: "ArcaneDoor is a campaign that is the latest example of state-sponsored actors targeting perimeter network devices"; "we identified a previously unknown actor now tracked as UAT4356 by Talos."; "we assess with high confidence that these actions were performed by a state-sponsored actor."; the two backdoors "Line Runner" and "Line Dancer"; "we identified actor-controlled infrastructure dating back to early November 2023"; "The investigation that followed identified additional victims, all of which involved government networks globally."; CVE-2024-20353 and CVE-2024-20359. the source ↗cited in The Lock on the List
- Cisco Security Advisories: cisco-sa-asaftd-webvpn-z5xP8EUB (CVE-2025-20333, Critical, CVSS 9.9) and cisco-sa-asaftd-webvpn-YROOTUW (CVE-2025-20362), first published 25 September 2025 at 16:00 GMT, fixed releases only, no workarounds; cisco-sa-asaftd-persist-CISAED25-03, "Continued Evolution of Persistence Mechanism Against Cisco Secure Firewall Adaptive Security Appliance and Secure Firewall Threat Defense", first published 23 April 2026: "the ArcaneDoor threat actor has developed a previously unknown persistence mechanism" residing in "the Cisco Firepower eXtensible Operating System (FXOS)". the source ↗cited in The Lock on the List
- Cisco Systems, Form 10-K for the fiscal year ended 25 July 2026, filed 2 September 2026, Management's Discussion and Analysis: revenue by product category, Networking $34,668 million, Security $8,232 million, Collaboration $4,300 million, Observability $1,095 million, Services $15,030 million, total $63,325 million; "Revenue in our Security product category increased by 2%, or $0.1 billion, primarily driven by growth in our SASE and Network Security offerings."; "Total revenue was $63.3 billion, an increase of 12% compared with fiscal 2025." the source ↗cited in The Lock on the List
- CNBC (Jonathan Vanian), 'Anthropic and OpenAI need truly independent safety evaluators, experts say in public letter', 18 September 2026 the source ↗cited in Where They Land
- AI Evaluator Forum, 'Minimum Conditions for Embedding Evaluators', public letter, published 18 September 2026, last updated 23 September 2026; the letter's own page the source ↗cited in Where They Land
- METR, 'About METR', the organisation's own page the source ↗cited in Where They Land
- METR, home page, 'Funding update', 14 August 2026 the source ↗cited in Where They Land
- NBC News (Jared Perlo), 'Two AI researchers leave Anthropic and Google over safety concerns', 10 September 2026 the source ↗cited in Where They Land
- The Register (Richard Speed), 'OpenAI tells employees it won't claw back their vested equity', 24 May 2024 the source ↗cited in The Silence Contracts
- 'A Right to Warn about Advanced Artificial Intelligence', the open letter's own page, dated 4 June 2024 the source ↗cited in The Silence Contracts
- California Legislature, SB 53 (Wiener), Chapter 138, Statutes of 2025, bill text as chaptered: Labor Code chapter 5.1, 'Whistleblower Protections: Catastrophic Risks in AI Foundation Models' the source ↗cited in The Silence Contracts
- NBC News (Jared Perlo), 'Two AI researchers leave Anthropic and Google over safety concerns: There are no adults in the room', 10 September 2026 the source ↗cited in The Resignation Letters
- 'A Right to Warn about Advanced Artificial Intelligence', the open letter's own page, dated 4 June 2024 the source ↗cited in The Resignation Letters
- U.S. Senate, Committee on the Judiciary, Subcommittee on Privacy, Technology, and the Law, hearing 'Oversight of AI: Insiders' Perspectives', 17 September 2024, S. Hrg. 118-779, the printed record from the U.S. Government Publishing Office (package CHRG-118shrg63050) the source ↗cited in The Resignation Letters
- William Saunders, written testimony to the U.S. Senate Judiciary Subcommittee on Privacy, Technology, and the Law, 17 September 2024 (PDF on the committee's site) the source ↗cited in The Resignation Letters
- CNBC (Hayden Field), 'OpenAI disbands another safety team, as head advisor for AGI Readiness resigns', 24 October 2024 the source ↗cited in The Resignation Letters
- Fortune, 'Another OpenAI researcher quits—claims AI labs are taking a very risky gamble with humanity amid the race toward AGI', 28 January 2025 the source ↗cited in The Resignation Letters
- BBC News (Liv McMahon and Ottilie Mitchell), 'AI safety leader says world is in peril and quits to study poetry', 13 February 2026 the source ↗cited in The Resignation Letters
- TechCrunch (Anthony Ha), 'OpenAI hardware exec Caitlin Kalinowski quits in response to Pentagon deal', 7 March 2026 the source ↗cited in The Resignation Letters
- Pacing the Frontier, the statement's own page the source ↗cited in AI: Pause for Safety
Humans
The people of the 2026 Replicants list who sit in Technology: 52, by rank. A card opens the person on the Replicants page.
#1Elon MuskChief executive of SpaceX (which now includes xAI) and TeslaModel labs
#2Sam AltmanChief executive, OpenAIModel labs
#4Mark ZuckerbergChairman and chief executive, MetaModel labs
#7Demis HassabisCo-founder and chair of Google DeepMind; chief scientist of AlphabetScience interlinkedModel labs
#8Dario AmodeiCo-founder and chief executive, AnthropicModel labs
#11Andrej KarpathyResearcher, pretraining team, AnthropicModel labs
#12Yann LeCunFounder and executive chairman, AMI Labs; professor, NYUScience interlinkedModel labs
#19Ilya SutskeverCo-founder and chief executive, Safe SuperintelligenceModel labs#21Liang WenfengFounder and chief executive, DeepSeekModel labs
#22Mira MuratiCo-founder and chief executive, Thinking Machines LabModel labs
#23Alexandr WangChief AI Officer, MetaModel labs
#26Mustafa SuleymanChief executive, Microsoft AIModel labs
#30Greg BrockmanCo-founder and president, OpenAIModel labs
#31Daniela AmodeiCo-founder, president and board chair, AnthropicModel labs
#35Arthur MenschCo-founder and chief executive, Mistral AIModel labs
#36Clem DelangueCo-founder and chief executive, Hugging FaceModel labs
#46Emad MostaqueFounder, Intelligent InternetModel labs
#47Aravind SrinivasCo-founder and chief executive, PerplexityModel labs
#48Jack ClarkCo-founder and head of public benefit, AnthropicPolitics interlinkedModel labs
#49Chris OlahCo-founder and interpretability research lead, AnthropicScience interlinkedModel labs
#50Jared KaplanCo-founder and chief science officer, AnthropicScience interlinkedModel labs
#52John SchulmanChief scientist, Thinking Machines LabScience interlinkedModel labs
#53Nat FriedmanHead of product, Meta Superintelligence LabsModel labs
#70Simon WillisonIndependent software engineer and writerContent producers
#71George Hotz (geohot)Founder, the tiny corp (tinygrad); president, comma.aiContent producers
#73Noam ShazeerAI scientist, OpenAIModel labs
#74Aidan GomezCo-founder and chief executive, CohereModel labs
#77Bret TaylorChair of the OpenAI board; co-founder and chief executive, SierraModel labs
#78Daniel GrossAI products, Meta Superintelligence LabsModel labs
#79Barret ZophCTO of Thinking Machines LabModel labs
#80Koray KavukcuogluSenior vice president, Google DeepMind; chief AI architect, GoogleModel labs
#81Shane LeggCo-founder and chief AGI scientist, Google DeepMindScience interlinkedModel labs#82Sam McCandlishCo-founder and chief architect, AnthropicModel labs
#83Igor BabuschkinCo-founder of xAIModel labs
#84Jan LeikeLead, Alignment Science, AnthropicScience interlinkedModel labs
#85Helen TonerInterim executive director, Center for Security and Emerging TechnologyPolitics interlinkedPolicy / Politics
#87Guillaume LampleCo-founder and chief scientist, Mistral AIScience interlinkedModel labs#88Timothée LacroixCo-founder and chief technology officer, Mistral AIModel labs
#89Fidji SimoPart-time adviser, OpenAI (former CEO of AGI Deployment / Applications)Model labs
#90Adam D'AngeloCo-founder and chief executive, QuoraModel labs
#113Shawn Wang (swyx)Co-founder and editor, Latent Space; founder, Smol AIContent producers
#114Andrew BosworthChief technology officer and head of Reality Labs, MetaIndustry interlinkedHyperscalers
#115Chris CoxChief product officer, MetaIndustry interlinkedHyperscalers
#118John GiannandreaFormer SVP, Machine Learning and AI Strategy, Apple (retired)Industry interlinkedHyperscalers
#119Craig FederighiSenior vice president, software engineering, AppleIndustry interlinkedHyperscalers
#156Nick CleggFormer President of Global Affairs, MetaPolitics interlinkedPolicy / Politics
#166Alessio FanelliCo-host, Latent Space; partner, DecibelContent producers#167Nathan LabenzHost, The Cognitive Revolution podcastContent producers#169Anna MakanjuHead of AI for Civil Society and Philanthropy, OpenAI FoundationPolitics interlinkedPolicy / Politics
#188Tim ScarfeHost and CEO, Machine Learning Street TalkContent producers#195Sam CharringtonHost, TWIML AI PodcastContent producers
#200Nathan LambertCo-founder and executive director, Trillium LabsContent producersVideos
OFF-WORLD · TECHNOLOGY, SAID OUT LOUDALL TECHNOLOGY ON OFF-WORLD →