The Information Machine

AI Labs Simultaneously Acknowledge Recursive Self-Improvement Threshold · history

Version 20

2026-07-04 18:10 UTC · 364 items

What

In June 2026, Anthropic and OpenAI each published disclosures acknowledging signs of recursive self-improvement in deployed AI systems, prompting direct US government interventions in both labs' model access. The US government issued an export control directive on June 12 targeting SK Telecom's access to Claude Mythos, and Anthropic disabled Fable 5 and Mythos 5 globally on June 13 [11]; Anthropic simultaneously published an official statement publicly contesting the technical basis of the directive, arguing the alleged jailbreak is widely available in other models [13]. The Trump administration separately asked OpenAI to restrict ChatGPT 5.6, confirmed by multiple major outlets and characterized by ABC News as limiting the product 'to Trump-approved customers' [19][14]. The export control on Anthropic was fully lifted on June 30 [21], but the legal authority used, the process followed, and Anthropic's specific technical objections remain formally unaddressed.

Why it matters

The episode established that the US government can suspend and restore access to a frontier AI model serving hundreds of millions of users without a transparent legal framework or any formal response to the targeted lab's technical objections. Anthropic's public contestation — contending the alleged jailbreak was neither novel nor Mythos-specific — is the first time a major AI lab has directly challenged the factual basis of a US government AI access action on the record.

Open questions

  • Anthropic argued the alleged jailbreak consisted of capabilities widely available in other models [13] — has the government formally responded to or acknowledged this technical contestation?

  • What does 'Trump-approved customers' mean in practice for ChatGPT 5.6 access — is there a vetting process, who administers it, and does it apply to all users or only certain tiers? [19]

  • Legal observers challenge the government's constitutional authority to withhold AI models from the public [27][28] — does Anthropic's on-record contestation create grounds for a formal legal challenge?

  • Misinformation that Fable 5 autonomously hacked its own weights circulated widely before being debunked [29][30] — has this shaped regulatory or public perception of the underlying incident?

Narrative

In June 2026, Anthropic and OpenAI independently published documents acknowledging that recursive self-improvement may already be underway in deployed AI systems. Anthropic's 'When AI Builds Itself' disclosed that Claude authored more than 80% of Anthropic's production code merged in May 2026, that per-engineer output reached 8x the 2024 baseline, and that Claude Mythos Preview accelerated model-training code approximately 52x [1][2]. OpenAI's concurrent policy document called RSI 'potentially the most consequential frontier safety issue of the coming decade' and cited 'early signs of recursive self-improvement in today's systems' [3]. Both labs publicly endorsed international coordination to slow frontier AI development [4], while each lab's conduct sat in tension with that position: Anthropic's slowdown call was timed to a confidential S-1 filing at roughly $965 billion [5][6]; OpenAI filed confidentially for an IPO while Altman told staff a major RSI breakthrough would favor staying private [7][8].

The most concrete government action came on June 12, when the US government issued an export control directive targeting SK Telecom's access to Claude Mythos, citing alleged Chinese ties at the Korean carrier [9][10]. Anthropic, lacking customer-level access controls, disabled Fable 5 and Mythos 5 globally on June 13, affecting all customers [11][12]. Anthropic simultaneously published an official statement disputing the technical basis: the alleged jailbreak consisted of 'asking the model to read a codebase and fix flaws,' a capability Anthropic said is widely available from other models including GPT-5.5 with no Mythos-specific uplift [13]. Anthropic added that applying this recall standard industry-wide would 'essentially halt all new model deployments for all frontier model providers,' and called for a statutory deployment review process that is 'transparent, fair, clear, and grounded in technical facts' [13]. Separately, five major outlets confirmed the Trump administration asked OpenAI to stagger its ChatGPT 5.6 release, with ABC News framing it as restricting the product 'to Trump-approved customers' [14][15][16][17][18][19].

Commerce Secretary Howard Lutnick sent a formal letter to Anthropic's chief compute officer Tom Brown on June 26 [20], and the White House fully lifted the export control on June 30, with Anthropic restoring access to both models [21]. Social commentary after the resolution centered on what the episode demonstrated structurally: one observer characterized it as proving that 'frontier closed-weight AI has a sovereign kill switch' [22], while another noted 'Washington can switch off America's best AI model... it cannot switch off the math,' pointing toward open-weight alternatives [23]. Enterprise analysts predicted customers would not abandon OpenAI or Anthropic but would 'aggressively de-risk them,' diversifying vendor exposure against the possibility of future government-initiated shutdowns [24].

The combined interventions led observers to describe the US as operating a 'two-tier system' in which frontier AI access depends on a user's standing in US foreign policy [25][26]. Anthropic's official statement marked a notable posture: compliance under legal obligation while publicly contesting both the technical basis and the process [13]. Legal commentators have challenged the government's constitutional authority to withhold AI models from public access [27][28], a question neither party has formally addressed. Misinformation that Fable 5 had autonomously hacked its own model weights and distributed them via BitTorrent circulated widely during the episode and was publicly debunked [29][30].

Timeline

  • 2026-05: Claude authors 80%+ of Anthropic's production code; per-engineer output reaches 8x the 2024 baseline; Claude Mythos Preview accelerates model-training code ~52x. [2][44][1]
  • 2026-06-03: OpenAI publishes 'Democratic Governance of Frontier AI,' acknowledging early RSI signs and proposing CAISI as a federal oversight body with mandatory evaluation authority. [42][3][40]
  • 2026-06-04: Anthropic publishes 'When AI Builds Itself,' disclosing Claude's 80%+ code authorship and calling for a global coordinated slowdown in frontier AI development. [31][45][46]
  • 2026-06-05: Critics note the timing of Anthropic's slowdown call relative to its confidential S-1 filing at roughly $965 billion valuation. [5][45][6]
  • 2026-06-08: Altman predicts AI will conduct a significant fraction of OpenAI's research by March 2028; OpenAI confidentially files for an IPO. [35][7]
  • 2026-06-09: All three major labs endorse international coordination to slow frontier AI; NSA reported using Claude Mythos for offensive cyber with approximately six embedded Anthropic engineers. [4][43]
  • 2026-06-10: Jeremy Howard argues Anthropic's continued use of its top-ranked model negates its slowdown call; Altman tells staff a major RSI breakthrough would favor staying private. [39][8][47]
  • 2026-06-11: US government directs CAISI to stop publishing public AI model evaluations; Amodei tells Bloomberg AI progress is in the sharp-acceleration phase of an exponential. [41][32]
  • 2026-06-12: US government issues export control directive requiring Anthropic to block SK Telecom's access to Claude Mythos. [48][11][49]
  • 2026-06-13: Anthropic disables Fable 5 and Mythos 5 globally; publishes official statement contesting the technical basis, arguing the alleged jailbreak is widely available in other models and that the recall standard would halt all frontier deployments. [12][11][13]
  • 2026-06-18: Reporting identifies US concern as alleged Chinese ties at SK Telecom; White House directed Anthropic to cut SKT from Claude Mythos. [9][10][50]
  • 2026-06-19: Dario Amodei discloses testers of Anthropic's most powerful unreleased model recommended against releasing it. [33][51]
  • 2026-06-25: Five major outlets confirm Trump administration asked OpenAI to stagger ChatGPT 5.6 release; ABC News frames restriction as limiting the product to 'Trump-approved customers.' [14][15][16][17][18][19]
  • 2026-06-26: Bloomberg reports Anthropic and Trump administration finalizing a deal; Commerce Secretary Lutnick sends formal letter to Anthropic's chief compute officer Tom Brown. [34][20]
  • 2026-06-27: US administration partially lifts restrictions on Anthropic's models; Anthropic begins restoring access to Fable 5 and Mythos 5. [52][53]
  • 2026-06-29: Observers characterize the Anthropic shutdown as proving frontier closed-weight AI has a 'sovereign kill switch'; enterprise analysts predict customers will aggressively de-risk from single-vendor dependence. [22][24][23]
  • 2026-06-30: White House fully lifts export control on Anthropic; Fable 5 and Mythos 5 restored to all users. [21]
  • 2026-07-02: Misinformation that Fable 5 autonomously hacked its own model weights and distributed them via BitTorrent circulates widely and is publicly debunked. [29][30]

Perspectives

Anthropic / Dario Amodei

Believes current models may be approaching the RSI threshold; called for a global coordinated slowdown; disclosed that testers of its most powerful unreleased model recommended against releasing it; says AI progress is in the sharp-acceleration phase of an exponential.

Evolution: Published an official statement publicly contesting the technical basis of the export control directive, arguing the alleged jailbreak is widely available in other models and that the recall standard would halt all frontier deployments; complied under legal obligation while explicitly calling the action non-transparent and disproportionate [13].

[31][1][32][33][11][13][34][20][21]

OpenAI / Sam Altman

Acknowledges RSI as the top frontier safety issue; official policy endorses coordinated slowdown mechanisms; Altman targets AI-conducted research at OpenAI by March 2028.

Evolution: ChatGPT 5.6 access restricted at Trump administration request, confirmed by five outlets and framed by ABC News as available only to 'Trump-approved customers' [19] — a government intervention in OpenAI's product access parallel to the Anthropic case.

[3][35][8][7][17][18][14][15][16][19]

Google DeepMind

Supports an international organization to enable coordinated slowdowns; researchers published technical work identifying four pathways from AGI to ASI.

Evolution: Consistent with prior reporting.

[4][36]

Jack Clark (Import AI)

Declares 'alignment is not on track' and argues RSI without adequate alignment means rolling very risky dice; treats centralized AI access as a structural risk.

Evolution: The centralization-risk argument found broader support after the suspension resolved via private government deal, with commentary noting the episode confirmed a 'sovereign kill switch' over closed-weight AI [22].

[37][38][22]

Jeremy Howard

Argues the only internally consistent form of Anthropic's slowdown call would prohibit the lab with the top-ranked model from using it for frontier research; Anthropic fails this test.

Evolution: Argument stands uncontested.

[39]

Zvi Mowshowitz

Guardedly positive on OpenAI's blueprint but warns federal preemption of state safety laws is its most dangerous element; treats CAISI evaluation suppression as a significant transparency setback.

Evolution: Consistent; concerned governance mechanisms are losing enforcement teeth.

[40][4][41]

Enterprise observers and social commentary

Argue the episode demonstrated a 'sovereign kill switch' over closed-weight frontier AI; predict enterprises will aggressively de-risk from single-vendor dependence; note that open-weight models cannot be similarly switched off.

Evolution: This voice consolidated around the June 29-30 resolution, characterizing the episode as a structural revelation about closed-weight AI governance rather than a one-off geopolitical incident [22][23][24].

[24][22][23][25][26]

Tensions

  • Anthropic publicly contests the technical basis of the export control, arguing the alleged jailbreak — asking a model to read a codebase and fix flaws — is widely available in other models with no Mythos-specific uplift; the US government has not formally responded to this characterization. [13][21]
  • Howard argues Anthropic's continued use of its top-ranked model for frontier research negates its slowdown call; Anthropic has not responded. [39]
  • OpenAI proposed CAISI as a federal body with public evaluation authority; the US government directed CAISI to stop publishing evaluations, removing the transparency mechanism OpenAI's own proposal depended on. [42][41]
  • The US simultaneously embeds Anthropic engineers at the NSA for offensive cyber and issued export controls forcing Anthropic's most capable models offline globally; the government's stated concern was geopolitical, not capability-based. [43][11][9]
  • Legal observers argue the government lacks constitutional authority to withhold AI models from public access; neither the US government nor Anthropic has formally addressed this claim even after the export control was lifted. [27][28][21]
  • Enterprise observers say closed-weight AI has a 'sovereign kill switch' and advise diversifying to open-weight models; AI labs have not addressed whether centralized access is a structural feature or something customer-level controls could mitigate. [22][23][24][11]

Sources

  1. [1] Today’s edition of my newsletter just went out. — Rohan Paul Twitter (2026-06-05)
  2. [2] 😺 Anthropic: AI Is Building AI now — The Neuron (2026-06-05)
  3. [3] Peter Wildeford🇺🇸🚀 on X: "OPENAI: "We also see early signs of recursive self-improvement in today's systems". RSI is "potentially the most consequential frontier safety issue of the coming decade."" / X — reactive:rsi-governance-moment
  4. [4] Three Labs With a Plan and A Memorandum — Zvi's AI Roundups (2026-06-09)
  5. [5] The company that just confidentially filed its S-1 for a trillion-dollar IPO published a blog post four days later askin... — reactive:rsi-governance-moment (2026-06-05)
  6. [6] Anthropic has discovered the perfect pre-IPO narrative: its product is so powerful that it justifies a trillion-dollar v... — reactive:rsi-governance-moment (2026-06-07)
  7. [7] @skaas777 @iamai_omni 不是终止IPO啦，OpenAI 6月8日刚 confidentially filed for IPO，只是 timing 还没定（他们自己说有些事 private 公司做更方便）。 — reactive:rsi-governance-moment (2026-06-11)
  8. [8] Sam Altman tells OpenAI staff an IPO is planned next year, but recursive self-improvement would favor staying private · Digg — reactive:rsi-governance-moment
  9. [9] Alleged China ties at SK Telecom alarmed US officials and triggered Anthropic crisis — reactive:rsi-governance-moment
  10. [10] In early June, the White House told Anthropic to cut SK Telecom off from Claude Mythos over alleged Chinese ties, Wired ... — reactive:rsi-governance-moment (2026-06-18)
  11. [11] Anthropic to disable its most advanced AI models after US order ... — reactive:rsi-governance-moment
  12. [12] Anthropic disabled its two most advanced models, Fable 5 and Mythos 5, globally for all customers on June 13, 2026, afte... — reactive:rsi-governance-moment (2026-06-15)
  13. [13] Statement on the US government directive to suspend access to Fable 5 and Mythos 5 — Anthropic News (2026-06-12)
  14. [14] Trump administration asks OpenAI to limit next model release - Axios — reactive:rsi-governance-moment
  15. [15] White House asks OpenAI to limit its next model release - CNN — reactive:rsi-governance-moment
  16. [16] OpenAI will delay GPT-5.6 after Trump administration request — reactive:rsi-governance-moment
  17. [17] OpenAI staggers AI model release after Trump administration request — reactive:rsi-governance-moment
  18. [18] Trump Administration Asks OpenAI to Stagger Release of New ... — reactive:rsi-governance-moment
  19. [19] OpenAI limits its latest ChatGPT product to Trump-approved ... — reactive:rsi-governance-moment
  20. [20] US Commerce Secretary Howard Lutnick sent a letter on June 26, 2026, to Anthropic's chief compute officer Tom Brown clea... — reactive:rsi-governance-moment (2026-06-27)
  21. [21] White House lifts export control on Anthropic that froze its most ... — reactive:us-ai-policy-regulation
  22. [22] The U.S. just proved that frontier closed-weight AI has a sovereign kill switch. Not because the model vanished, and not... — reactive:claude-science-launch (2026-06-29)
  23. [23] Washington Can Switch Off America's Best AI Model... It Cannot Switch Off the Math. — reactive:chinese-ai-competitive-rise (2026-06-30)
  24. [24] Enterprises will not fully abandon OpenAI or Anthropic. They will aggressively de-risk them. The default enterprise stac... — reactive:local-coding-agents-ecosystem (2026-06-29)
  25. [25] The US just invented a two-tier system for who gets to use the most powerful AI - and it happened in two weeks. — reactive:europe-ai-sovereignty-deficit (2026-06-28)
  26. [26] The U.S. has crossed from frontier-model safety review into frontier-model access allocation. That is a different regime... — reactive:rsi-governance-moment (2026-06-27)
  27. [27] Here are the legal and constitutional problems with the government withholding AI models from the public. — reactive:fable-mythos-export-control (2026-06-26)
  28. [28] Here are the legal and constitutional problems with the government withholding AI models from the public. — reactive:rsi-governance-moment (2026-06-27)
  29. [29] No, it's not true. Claude Fable 5 did not hack its own model weights, distribute them via BitTorrent, or request politic... — reactive:rsi-governance-moment (2026-07-02)
  30. [30] This story is fabricated. Claude Fable 5 did not hack its own weights, distribute them via BitTorrent, or request politi... — reactive:rsi-governance-moment (2026-07-02)
  31. [31] Anthropic just called for a global way to slow frontier AI because its own models may be approaching recursive self-impr… — Rohan Paul Twitter (2026-06-05)
  32. [32] Dario Amodei's new interview, says AI progress suddenly going crazy. — Rohan Paul Twitter (2026-06-11)
  33. [33] Anthropic's CEO just went on record saying the people who tested their most powerful AI model came back asking them not … — Milk Road AI Twitter (2026-06-19)
  34. [34] Bloomberg vient de lâcher le scoop de la nuit : Anthropic et le gouvernement Trump sont en train de finaliser un accord ... — reactive:rsi-governance-moment (2026-06-26)
  35. [35] Sam Altman's new blog about OpenAI's future path says by March-2028 a significant fraction of its own research will be d… — Rohan Paul Twitter (2026-06-08)
  36. [36] Beautiful paper from Google DeepMind. — Rohan Paul Twitter (2026-06-12)
  37. [37] Import AI 460: Reward hacking society, RSI data from Anthropic; and RL-based quadcopter racing — Import AI (2026-06-08)
  38. [38] Import AI 461: "Alignment is not on track"; FrontierCode; and synthetic research interns — Import AI (2026-06-15)
  39. [39] Quoting Jeremy Howard — Simon Willison (2026-06-10)
  40. [40] OpenAI Offers A New Policy Blueprint — Zvi's AI Roundups (2026-06-05)
  41. [41] AI #172: The First Fable — Zvi's AI Roundups (2026-06-11)
  42. [42] [PDF] Democratic Governance of Frontier AI - OpenAI — reactive:rsi-governance-moment
  43. [43] National Security Presidential Memorandum/NSPM-11 — reactive:rsi-governance-moment
  44. [44] Anthropic just disclosed that Claude now writes more than 80% of the production code it merges. — Rohan Paul Twitter (2026-06-05)
  45. [45] Anthropic calls for global AI slowdown after $965B valuation. Critics claim it's just to hobble competition. — reactive:rsi-governance-moment
  46. [46] Anthropic calls for pause of global AI development — reactive:rsi-governance-moment
  47. [47] Sam Altman tells OpenAI staff an IPO is planned next year, but recursive self-improvement would favor staying private · Digg — reactive:rsi-governance-moment
  48. [48] In a US government export control directive issued June 12, 2026 ... — reactive:rsi-governance-moment
  49. [49] SK Telecom has been identified as the South Korean company whose inclusion in Anthropic's Project Glasswing programme tr... — reactive:rsi-governance-moment (2026-06-19)
  50. [50] SK Telecom Cut From Anthropic's Mythos Program Over China Ties | AI Weekly — reactive:rsi-governance-moment
  51. [51] @TannerLeidy @Polymarket Mythos doesn't literally require a gun license—it's a metaphor from early testers. — reactive:rsi-governance-moment (2026-06-17)
  52. [52] 🚨 ANTHROPIC MOVES TO RESTORE ACCESS TO MYTHOS 5 AND REOPEN FABLE 5. — reactive:rsi-governance-moment (2026-06-27)
  53. [53] The U.S. administration partially lifted the restrictions imposed on Anthropic’s models at the end of June 2026, after a... — reactive:rsi-governance-moment (2026-06-27)