Self-regulation by AI companies is inherently limited by conflict of interest. Discuss with reference to recent frontier-AI misuse disclosures.

Frontier AI is dual-use: the same firms that commercialise these models are also their principal policing authority. Anthropic's fourth threat-intelligence report, Countering misuse of AI (September 2026), disrupting misuse across seven harm categories between December 2025 and August 2026 [1], is the latest illustration of both the value and the structural limits of this arrangement.

Where the conflict of interest bites

• The reporter is the adjudicator: the harm taxonomy, severity grading and claim of disruption all originate with the developer, with no external verification step [1].
• The denominator is withheld: disruptions are disclosed, but not attempt volumes, success rates or time-to-detection [1] — a firm choosing the metrics that flatter it.
• Safety concessions arrive late: weaker safeguards in superseded model versions are conceded only once admitting them is commercially costless [1].
• Safeguards are reactive: pattern-based classifiers act after misuse accumulates, so detection is not prevention — a lag that agentic, multi-step attack chains exploit [1].

What self-disclosure still delivers

• Only the developer holds model internals and logs; identifying state-sponsored groups, commercial spyware vendors and financially motivated criminals as actual users [1] yields threat intelligence no treaty process currently generates.

Why external oversight has not closed the gap

• UN High Commissioner Volker Türk calls voluntary self-regulation "nowhere near sufficient," urging mandatory incident reporting, capability verification and human-rights due diligence [2].
• The Global Dialogue on AI Governance (UNGA Resolution A/RES/79/325, 2025) ends in co-chair summaries, not binding instruments [3] — norms without an inspectorate.
• India's AI Governance Guidelines adopt a principle-based techno-legal route with an AI Safety Institute and AI Governance Group [4], embedding controls into system design [5], yet upstream training decisions lie outside domestic jurisdiction.

Self-regulation is therefore necessary but insufficient: the asymmetry of capability justifies the developer as reporter, never as judge. The way forward is standardised, auditable disclosure metrics and statutory telemetry access for empowered safety institutes, so that accountability — as India's guidelines affirm — is engineered in rather than volunteered.