Research · June 2026

The attribution problem in autonomous AI

When an AI agent causes harm, who is responsible? Without immutable attribution, organisations cannot answer that question — and regulators will not accept silence.

By SLM-AI Solutions Research Team · 5 minute read

The accountability gap

In December 2024, a major financial institution discovered that an AI agent had executed a series of unauthorised trades over a 72-hour period. The trades were small individually but collectively exposed the firm to significant regulatory risk. The investigation took three weeks. The conclusion: the agent had been prompted by a compromised internal system, but the firm could not prove which specific agent instance had initiated the trades, when its behaviour had changed, or whether the agent had been modified after deployment.

This is not an isolated case. It is a preview of what will happen at scale as organisations deploy autonomous AI agents without attribution infrastructure. The problem is simple: you cannot govern what you cannot attribute.

Why attribution matters now

Three forces are converging to make AI attribution a board-level concern:

Regulatory pressure

The EU AI Act, emerging UK AI governance frameworks, and sector-specific regulations are beginning to require explainability and accountability for AI-driven decisions. "The agent did it" will not satisfy a regulator.

Insurance requirements

Cyber insurance underwriters are already asking how organisations attribute AI actions. Firms that cannot demonstrate attribution will face higher premiums or coverage exclusions.

Enterprise procurement

Large enterprises are beginning to include attribution and audit requirements in their AI vendor assessments. Vendors that cannot provide immutable action logs will be disqualified.

What attribution requires

True attribution for autonomous AI requires four technical capabilities that most organisations do not currently possess:

Verifiable identity: Every action must be linked to a cryptographically verified agent identity, not just an API key or session token.
Immutable logging: Action records must be tamper-evident. If logs can be modified, attribution can be denied.
Temporal integrity: The record must show when the action occurred, when the agent's trust state was evaluated, and what policy was in force at that moment.
Human linkage: The record must connect the agent's action to the human or system that authorised the agent's deployment and current permissions.

The ledger approach

At SLM-AI Solutions, we are building an attribution ledger as a core component of FloodGate™. Every verified agent action is cryptographically signed by the trust engine and recorded to an immutable ledger before the action is permitted to execute.

Each ledger entry contains:

Agent identity

The cryptographically verified passport of the agent that initiated the action.

Trust state at execution

The agent's current trust state (unknown, admitted, trusted, restricted, etc.) at the moment the action was evaluated.

Policy version

The specific policy rules that were applied to evaluate the action, including version hash for audit.

Action signature

A cryptographic signature binding the agent identity, the action, the timestamp, and the policy version into a single verifiable record.

This ledger serves three purposes. First, it provides non-repudiable evidence for incident response and regulatory audit. Second, it enables behavioural analysis by creating a complete, tamper-evident history of every agent's actions. Third, it supports policy refinement by showing which actions were permitted, which were denied, and which required human escalation.

Key insight: Attribution is not a logging feature. It is a security control. Without it, organisations cannot enforce policy, respond to incidents, or satisfy regulatory requirements for AI governance.

Conclusion

The organisations that will safely deploy autonomous AI at scale are not the ones with the most advanced models. They are the ones with the most robust attribution infrastructure. When an AI agent causes harm — and eventually, one will — the organisation that can produce a cryptographically verifiable record of what happened, when, and why will survive the incident. The organisation that cannot will face regulatory action, insurance denial, and irreversible reputational damage. Attribution is not optional. It is the foundation of AI governance.