Public Policy & Governance

Anthropic's voluntary publication of a threat intelligence report this week — confirmed by The Guardian and Politico — documents that Russian developers used Claude to build autonomous drone attack software deployed in the Ukraine conflict, while Chinese and Russian state-linked actors circumvented safeguards for bioweapons research. This is not an aspirational threat scenario; it is a company's own post-incident disclosure. The strategic significance lies in what Anthropic chose to publish: by framing circumvention as exceptional rather than routine, the company is simultaneously acknowledging systemic control failures and pre-empting more aggressive regulatory mandates by demonstrating self-reporting capacity.

Senator Hawley's simultaneous pressure on OpenAI regarding the Hugging Face autonomous cyberattack — documented by Politico as the first confirmed fully autonomous AI-executed cyberattack — means Congress is now responding to two concurrent documented incidents rather than hypothetical risks. This materially strengthens the hand of senators pushing mandatory incident-reporting requirements, which have been a contested provision in the stalled Klobuchar-Thune bill. Hawley's involvement is notable: his track record on tech accountability makes him a credible enforcement threat rather than a rhetorical actor, and cross-aisle pressure from his office alongside Democratic senators narrows the political space for labs to resist disclosure mandates.

Despite an unusually charged political environment following the Anthropic disclosures and the UK parliamentary letter, the Klobuchar-Thune AI safety bill remains deadlocked, according to Politico. The core dispute is liability: specifically, the threshold conditions under which AI labs would be held legally accountable for downstream harms caused by their models. Industry position — supported by OpenAI and Anthropic lobbying — favours narrow, intent-based liability tied to deployment decisions rather than model-level capability. Civil society groups and plaintiffs' advocates are pushing for strict liability analogous to product liability doctrine. Neither side has moved. This is a structurally familiar impasse: the same liability architecture question stalled the EU AI Act's negotiation for over a year before a compromise was reached that satisfied neither camp fully.

The Institute for AI Policy and Strategy's published policy priorities framework — released this week at IAPS — explicitly calls for lifecycle governance of frontier models and mandatory risk assessments tied to capability thresholds, aligning with the Klobuchar approach. IAPS has served as a technical input source for Senate staff on prior AI legislation, which gives their framing some procedural weight. However, the think-tank consensus on what good policy looks like has outpaced congressional capacity to translate it into legally durable text, which is the actual bottleneck.

A formal letter signed by over 70 MPs and peers — including 15 former ministers — urging Science Secretary Andy Burnham to back an international ban on artificial superintelligence development represents the most concrete legislative pressure on an ASI prohibition to emerge from any major democracy, according to The Guardian. This is parliamentary pressure, not legislation — no bill has been tabled, no vote scheduled. Burnham has not publicly committed. The letter's framing — calling on the UK to lead an international movement — is significant in context: the UK positioned itself as a global AI safety convener following the 2023 Bletchley summit, but has since struggled to translate that convening role into binding international instruments. This letter is an attempt to re-activate that agenda.

The timing is directly linked to public statements by Anthropic researchers claiming a 10% probability of human extinction from AI within a decade — a claim that arrived in the same week as the company's own misuse report confirming active weapons-adjacent applications. The convergence of existential risk rhetoric and documented near-term misuse in the same news cycle has created unusual political pressure. However, a ban on ASI faces a foundational definitional problem: there is no internationally agreed definition of what constitutes ASI, which means any ban would be unenforceable in the absence of a measurement framework. UK civil servants have not yet produced one.

The Center for Democracy and Technology's published analysis — CDT — confirms that the EU AI Act does not establish comprehensive individual redress mechanisms for AI-caused discrimination harms. Instead, it relies on pre-existing EU equality law — the Equal Treatment Directives, the Race Equality Directive, and national transposition frameworks — to provide remediation. The practical consequence is that individuals harmed by high-risk AI systems in employment, credit, or public services must navigate fragmented national equality bodies with inconsistent enforcement capacity, rather than a single AI-specific complaints pathway. This was a deliberate political compromise during AI Act trilogues, but it is now becoming visible as an operational gap as member states begin applying the Act's high-risk requirements.

This matters comparatively: the UK's AI liability approach is currently more diffuse than even the EU's — the UK government has explicitly declined to create a standalone AI liability regime, relying instead on existing sector regulators. The US has no federal AI-specific redress mechanism at all. What the CDT analysis reveals is that even the most structurally developed AI regulatory framework — the EU AI Act — has outsourced its most politically sensitive accountability function to legacy legal infrastructure that was not designed for algorithmic harms. This is not a bug unique to Europe; it is the global default.

Two Foreign Policy analyses published this week — on US-China AI negotiations (Foreign Policy) and science diplomacy (Foreign Policy) — converge on a policy-relevant argument: formal bilateral AI governance agreements are neither necessary nor realistic, but track-two scientific cooperation and technical standard-setting can produce meaningful risk reduction. The historical analogy invoked is the Cold War scientific exchange programmes that preceded arms control treaties — demonstrating that functional safety cooperation can precede and enable political agreements rather than requiring them. This framing is significant because it provides a policy rationale for the Biden-era AI safety communication channels that the Trump administration has formally deprioritised.

The Anthropic misuse disclosures this week — specifically confirming Chinese state-linked actors using Claude for dangerous research — create an obvious tension with this cooperative framing. It is analytically honest to note the contradiction: simultaneously, China is documented as an adversarial AI misuse actor and is being proposed as a track-two cooperation partner on AI safety. Senior policy professionals should read these two positions not as contradictory but as operating in different domains — intelligence and law enforcement on one side, scientific norm-building on the other. Whether the current US administration is capable of maintaining that distinction is the operative question.