AI self-development at Anthropic — May 2026
Key metrics from the Anthropic Institute recursive self-improvement report
Anthropic Brief: Dario calls for government authority to block dangerous AI — backed by data showing Claude now writes 80%+ of Anthropic's code
On June 10–11, Anthropic published its most aggressive governance stance yet: a dual policy framework proposing binding government authority to block frontier model deployments, paired with internal data showing Claude already authors more than 80% of Anthropic's code and delivers ~52× research speedups. This brief covers what Anthropic is proposing, the thresholds that would trigger oversight, and why the internal data release matters for enterprise and policy readers.
Anthropic published two closely linked documents on June 10–11 that mark its most aggressive public stance on AI governance to date: a dual policy framework calling for binding government authority over frontier models, and an internal data report showing AI systems are already writing the majority of the company's own code.
What Anthropic is asking governments to do
The "Policy on the AI Exponential" package, released June 10, contains two separate proposals.1
The Advanced AI Framework calls for mandatory pre-release testing, independent third-party audits, and enforceable government authority to block or reverse the deployment of any frontier model that poses a significant risk of catastrophic harm. Anthropic proposes the rules apply to any model trained above 10²⁵ floating-point operations (FLOPs) at companies earning more than $500 million in AI-related revenue or spending more than $1 billion on AI R&D. Civil penalties would be tied to global annual revenue and escalate with repeated violations.1
Four risk categories drive the proposal: biological weapons uplift, large-scale cyberattack enablement, loss of human control over AI systems, and AI systems autonomously accelerating their own development. Anthropic explicitly argues that transparency requirements alone — such as those already passed in California and New York — are no longer sufficient.
The Economic Policy Framework, the second document in the package, addresses labor market disruption. It calls for wage insurance, retention tax incentives for employers, a strengthened social safety net, and potential universal basic income or capital accounts. Dario Amodei also explicitly endorses reform at agencies like the FDA to accelerate approval of AI-discovered drugs, and calls for a ban on domestically deployed fully autonomous weapons.2
The framework is written primarily for the U.S. federal government, but Anthropic states it does not support federal preemption of state law unless any federal legislation meets or exceeds the standard it is proposing.
The internal data behind the ask
Published the following day, the Anthropic Institute's "When AI Builds Itself" report provides the empirical basis for the urgency.3
통계 카드를 불러오는 중…
The report describes three possible futures: capability stalls at current levels (least likely, per Anthropic), AI labs see compounding efficiency gains without full autonomy, or full recursive self-improvement begins. Anthropic states it does not believe the first scenario is likely.
Why it matters for enterprise and policy readers
The timing is deliberate. Claude Fable 5 and Mythos 5 shipped June 9 — the same week Anthropic filed a confidential S-1 and closed its $65B Series H.4 The policy push follows that capital event by 48 hours.
Anthropic's proposal is structurally self-applying: the company's own revenue almost certainly exceeds the $500M threshold it sets, and its training runs almost certainly exceed 10²⁵ FLOPs. Publishing the framework while also disclosing internal data on accelerating autonomous AI development is an unusual combination — it simultaneously positions Anthropic as responsible steward and first-mover advocate for the regulatory regime it would itself face.
For enterprise buyers and partners, the key near-term implication is potential mandatory third-party audits for models in their stack. If the framework were enacted, any vendor using a covered model would face indirect compliance questions about the models they deploy. The framework does not yet carry any enforcement weight, but signals the direction Anthropic plans to lobby.
| Framework element | Threshold / detail |
|---|---|
| Compute scope | > 10²⁵ FLOPs |
| Revenue scope | > $500M AI revenue or > $1B AI R&D |
| Penalties | Civil fines, % of global annual revenue, escalating |
| Required actions | Pre-release testing, system cards, risk reports, independent evaluation, security program |
| Government authority | Block or deter deployment of high-risk models |
이 콘텐츠를 둘러싼 관점이나 맥락을 계속 보강해 보세요.