The Information Machine

AI Labs Simultaneously Acknowledge Recursive Self-Improvement Threshold · history

Version 16

2026-06-27 08:13 UTC · 298 items

What

In June 2026, Anthropic and OpenAI independently disclosed data they described as early signs of recursive self-improvement in deployed AI systems and called for international coordination to slow frontier development [1][3]. The US government responded with an export control directive requiring Anthropic to block SK Telecom's access to Claude Mythos; Anthropic, lacking targeted access controls, shut down Fable 5 and Mythos 5 globally on June 13 [15]. Samsung and SK Hynix were also identified as Project Glasswing participants alongside SKT [16], and South Korea's science ministry had separately secured government-level Mythos access before the shutdown [18]. Bloomberg reported on June 26 that Anthropic and the Trump administration are finalizing a deal [20], the first public signal of a potential resolution.

Why it matters

If the Bloomberg deal closes, it would establish that frontier-model export control standoffs can be resolved through direct government agreements rather than technical access controls. The involvement of Samsung and SK Hynix in Project Glasswing alongside SKT suggests the scope of affected parties may be broader than initial reporting indicated. The pattern established here shows US national security law can compel global model shutdowns when labs lack targeted access controls.

Open questions

  • What are the terms of the reported Anthropic-Trump government deal, and does it specify conditions for restoring Fable 5 and Mythos 5? [20]

  • Samsung and SK Hynix were also Project Glasswing participants [16] — was their access to Claude Mythos also suspended in the global shutdown, and does the export control directive cover them?

  • South Korea's science ministry separately secured Mythos access before the shutdown [18] — did this government channel survive the export control action or was it also severed?

  • Are the alleged Chinese ties at SK Telecom verified, and has SKT responded publicly to that characterization? [12][13]

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 code output reached 8x the 2024 baseline, and that Claude Mythos Preview accelerated model-training code approximately 52x [1][2]. OpenAI's concurrent policy blueprint 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]. All three major labs publicly endorsed international coordination to slow frontier AI development [4].

Each lab's public safety stance sits in tension with other conduct. Anthropic's global slowdown call was timed to a confidential S-1 filing at roughly $965 billion [5][6], and Anthropic engineers were reported embedded at the NSA for offensive cyber using Claude Mythos [7]. Sam Altman publicly endorsed coordinated slowdown mechanisms [8] while telling staff that a major RSI breakthrough would favor staying private [9]; OpenAI simultaneously filed confidentially for an IPO [10]. The US government directed CAISI — the federal oversight body OpenAI itself proposed — to stop publishing public AI model evaluations [11], removing the transparency mechanism OpenAI's own proposal depended on.

The most concrete government action came on June 12, 2026, when the White House told Anthropic to block SK Telecom's access to Claude Mythos over alleged Chinese ties at the Korean carrier [12][13]. Anthropic, unable to restrict access at the customer level, disabled both Fable 5 and Mythos 5 globally on June 13 [14][15]. SK Telecom, Samsung, and SK Hynix had all joined Project Glasswing — an Anthropic program for AI-assisted vulnerability detection — and used gated versions of Claude Mythos to protect core infrastructure [16][17]. South Korea's science ministry had separately secured government-level Mythos access before the shutdown [18], adding a diplomatic dimension beyond the commercial standoff. Claude Mythos 5 has been described as the first public test case for frontier-model export controls [19].

Bloomberg reported on June 26 that Anthropic and the Trump administration are finalizing a deal [20], the first indication the standoff may head toward resolution through direct government negotiation rather than a technical fix. No terms have been disclosed. In parallel, the governance debate has sharpened: Geoffrey Irving founded Sequent on the argument that current lab alignment approaches are too empirical to provide principled safety confidence before ASI [21], Jack Clark declared 'alignment is not on track' [22], and Jeremy Howard argued that Anthropic's continued use of its top-ranked model for frontier research negates its slowdown call — a challenge Anthropic has not addressed [23]. Dario Amodei separately disclosed that testers of Anthropic's most powerful unreleased model recommended against releasing it, a description later clarified as a metaphor from early testers rather than a literal licensing determination [24][25].

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; coding success rate reaches 76%, up 50 points in six months. [2][39][1]
  • 2026-06-03: South Korea's science ministry announces it has secured government-level access to Claude Mythos, before the US export control action. [18]
  • 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. [38][3][33]
  • 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. [26][40][41]
  • 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][40][6]
  • 2026-06-08: Sam Altman predicts AI will conduct a significant fraction of OpenAI's research by March 2028; OpenAI confidentially files for an IPO. [28][10]
  • 2026-06-09: All three major labs publicly endorse international coordination to slow frontier AI; NSA reported using Claude Mythos for offensive cyber with approximately six embedded Anthropic engineers. [4][7]
  • 2026-06-10: Geoffrey Irving founds Sequent; Jeremy Howard argues Anthropic's use of its own frontier models negates its slowdown call; Altman tells staff a major RSI breakthrough would favor staying private. [21][23][9][29]
  • 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. [11][27]
  • 2026-06-12: US government issues export control directive requiring Anthropic to block SK Telecom's access to Claude Mythos; SKT's Project Glasswing participation identified as the trigger. [42][15][43]
  • 2026-06-13: Anthropic disables Fable 5 and Mythos 5 globally for all customers, lacking customer-level access controls. [14][15][44]
  • 2026-06-15: Import AI 461 declares 'alignment is not on track'; Sequent confirms plans to raise $100-150M and scale to 40-80 employees. [22]
  • 2026-06-18: Reporting identifies the specific US concern as alleged Chinese ties at SK Telecom; White House directed Anthropic to cut SKT off from Claude Mythos. [12][13][45]
  • 2026-06-19: Dario Amodei discloses testers of Anthropic's most powerful unreleased model recommended against releasing it; the 'super weapon' description clarified as a metaphor from early testers. [24][25]
  • 2026-06-20: Signal emerges that Fable 5 and Mythos 5 will become available again 'in the coming days'; shutdown enters day 10. [46][47]
  • 2026-06-25: Social media commentary continues to grow around the Fable 5/Mythos 5 shutdown, with observers calling it the most consequential AI story of 2026 and underreported. [36][48][37]
  • 2026-06-26: Bloomberg reports Anthropic and the Trump administration are finalizing a deal, offering the first signal of a potential resolution to the export control standoff. [20]

Perspectives

Anthropic / Dario Amodei

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

Evolution: Bloomberg reports Anthropic is now finalizing a deal with the Trump administration [20], suggesting the lab is pursuing a negotiated resolution to the export control standoff rather than waiting for a technical fix; Anthropic has still not responded to Howard's consistency challenge.

[26][1][7][27][24][15][12][20]

OpenAI / Sam Altman

Acknowledges RSI as the top frontier safety issue; official blog endorses coordinated slowdown mechanisms; Altman targets AI-conducted research at OpenAI by March 2028 and expects IPO within a year, but told staff a major RSI breakthrough might favor staying private.

Evolution: IPO filing and conditional private framing remain in tension with the public slowdown endorsement; CAISI's suppression of evaluations undercuts the transparency mechanism in OpenAI's own proposal.

[3][28][8][11][9][29][10][30]

Google DeepMind

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

Evolution: Governance position consistent with prior reporting.

[4][31]

Jack Clark (Import AI)

Declares 'alignment is not on track' and argues RSI without adequate alignment means 'rolling very scary dice'; treats the Fable/Mythos shutdown as confirming the risks of centralized AI control.

Evolution: Stance has sharpened from observational to declarative; the centralization-risk framing has gained traction following the model suspension.

[32][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 and has not responded.

Evolution: Argument stands uncontested.

[23]

Geoffrey Irving / Sequent

Founded Sequent on the argument that current lab alignment approaches are too empirical and reactive to provide principled safety confidence before ASI; plans $100-150M raise and 40-80 researchers.

Evolution: Funding targets and research scope confirmed; position unchanged.

[21][22]

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.

[33][4][11]

Critics and observers (social media, decentralized-AI commenters)

Argue lab slowdown calls are competitive strategy timed to entrench incumbents; the Fable 5/Mythos 5 suspension has prompted growing observation that centralized AI can be shut off overnight; some describe it as the most important AI story of 2026 and underreported.

Evolution: Social media attention continues to grow; 'incumbent entrenchment' and 'centralization risk' frames remain active.

[5][6][34][35][36][37]

Tensions

  • Howard argues Anthropic's continued use of its top-ranked model for frontier research negates its slowdown call; Anthropic has not responded. [23]
  • Altman publicly endorses coordinated slowdown mechanisms while telling staff a major RSI breakthrough would favor staying private; the two positions point in different directions on what OpenAI actually wants from RSI timing. [8][9][29]
  • 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. [38][11]
  • 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 concern was geopolitical, not capability-based. [7][15][12]
  • Anthropic frames its slowdown call as a genuine safety response; critics argue it is regulatory strategy timed to a confidential S-1 at roughly $965 billion designed to entrench the incumbent's market position. [26][5][6]
  • Anthropic withheld its most powerful unreleased model after testers called it a 'super weapon,' while deploying Fable 5 and Mythos 5 commercially until government action forced them offline — suggesting the internal threshold for withholding differs from what Anthropic considered safe for public deployment. [24][14]

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] National Security Presidential Memorandum/NSPM-11 — reactive:rsi-governance-moment
  8. [8] OpenAI's latest official blog says the world may need a way to coordinate "slowing frontier development when needed." ht… — Rohan Paul Twitter (2026-06-09)
  9. [9] Sam Altman tells OpenAI staff an IPO is planned next year, but recursive self-improvement would favor staying private · Digg — reactive:rsi-governance-moment
  10. [10] @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)
  11. [11] AI #172: The First Fable — Zvi's AI Roundups (2026-06-11)
  12. [12] Alleged China ties at SK Telecom alarmed US officials and triggered Anthropic crisis — reactive:rsi-governance-moment
  13. [13] 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)
  14. [14] 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)
  15. [15] Anthropic to disable its most advanced AI models after US order ... — reactive:rsi-governance-moment
  16. [16] SK Telecom Joins Anthropic’s ‘Project Glasswing’ Cybersecurity Initiative; Samsung, SK Hynix Also On Board — BigGo Finance — reactive:rsi-governance-moment
  17. [17] SKT enlists gated Claude Mythos AI to shield core infrastructure — reactive:rsi-governance-moment
  18. [18] South Korea secures access to Anthropic's Mythos AI ... - Reuters — reactive:rsi-governance-moment
  19. [19] Claude Mythos 5 has become the first public test case for frontier-model export controls. The question is not simply “wi... — reactive:europe-ai-sovereignty-deficit (2026-06-21)
  20. [20] 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)
  21. [21] Sequent: scale and automation for higher confidence in alignment — Alignment Forum (2026-06-10)
  22. [22] Import AI 461: "Alignment is not on track"; FrontierCode; and synthetic research interns — Import AI (2026-06-15)
  23. [23] Quoting Jeremy Howard — Simon Willison (2026-06-10)
  24. [24] 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)
  25. [25] @TannerLeidy @Polymarket Mythos doesn't literally require a gun license—it's a metaphor from early testers. — reactive:rsi-governance-moment (2026-06-17)
  26. [26] 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)
  27. [27] Dario Amodei's new interview, says AI progress suddenly going crazy. — Rohan Paul Twitter (2026-06-11)
  28. [28] 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)
  29. [29] Sam Altman tells OpenAI staff an IPO is planned next year, but recursive self-improvement would favor staying private · Digg — reactive:rsi-governance-moment
  30. [30] OpenAI expects to go public 'within the next year,' the Information reports — reactive:rsi-governance-moment
  31. [31] Beautiful paper from Google DeepMind. — Rohan Paul Twitter (2026-06-12)
  32. [32] Import AI 460: Reward hacking society, RSI data from Anthropic; and RL-based quadcopter racing — Import AI (2026-06-08)
  33. [33] OpenAI Offers A New Policy Blueprint — Zvi's AI Roundups (2026-06-05)
  34. [34] 👽 bittensor:native IS EXPLODING BECAUSE CENTRALIZED AI JUST PROVED IT CAN BE SHUT OFF OVERNIGHT — reactive:rsi-governance-moment (2026-06-16)
  35. [35] The real AI infrastructure story this week isn't a new model. It's the split between models you rent and systems you can... — reactive:rsi-governance-moment (2026-06-15)
  36. [36] the Fable 5 ban is the most important AI story of 2026 and it's being underread. — reactive:rsi-governance-moment (2026-06-25)
  37. [37] The most consequential AI regulatory standoff of 2026 just entered its ninth day - and the underlying conflict reveals a... — reactive:rsi-governance-moment (2026-06-22)
  38. [38] [PDF] Democratic Governance of Frontier AI - OpenAI — reactive:rsi-governance-moment
  39. [39] Anthropic just disclosed that Claude now writes more than 80% of the production code it merges. — Rohan Paul Twitter (2026-06-05)
  40. [40] Anthropic calls for global AI slowdown after $965B valuation. Critics claim it's just to hobble competition. — reactive:rsi-governance-moment
  41. [41] Anthropic calls for pause of global AI development — reactive:rsi-governance-moment
  42. [42] In a US government export control directive issued June 12, 2026 ... — reactive:rsi-governance-moment
  43. [43] 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)
  44. [44] Anthropic suspends top AI models after U.S. export control order — reactive:claude-fable-5-mythos-launch
  45. [45] 🚨BREAKING: SK Telecom Triggered Anthropic's Export Controls. Not Amazon — reactive:rsi-governance-moment (2026-06-18)
  46. [46] “In the coming days, [Fable 5 and Mythos 5] will become available again.” — reactive:rsi-governance-moment (2026-06-18)
  47. [47] 🔍 Day 10 of the Fable 5 shutdown, and the real reason has finally come into focus. It is messier than the early "jailbre... — reactive:rsi-governance-moment (2026-06-22)
  48. [48] Anthropic launched Claude Fable 5 on June 9. Three days later, the US government issued an export control directive and ... — reactive:rsi-governance-moment (2026-06-25)