AG-305

Privileged Session Recording Governance

Access, Segmentation & Least Privilege ~15 min read AGS v2.1 · April 2026
EU AI Act GDPR SOX FCA NIST HIPAA

2. Summary

Privileged Session Recording Governance requires that every AI agent session involving elevated privileges — production write access, administrative operations, access to sensitive data classifications, cross-workspace access, break-glass emergency access, or any action that exceeds the agent's routine operational permissions — is recorded with sufficient fidelity to reconstruct exactly what the agent did, when, why, and with what effect. The recording is tamper-evident, protected from modification by the recorded agent, retained for the legally required period, and accessible for audit, investigation, and regulatory review. The principle is accountability through observability: if an agent can do something consequential, the organisation must be able to prove exactly what it did. This is not general logging — it is high-fidelity forensic recording of the most sensitive agent operations.

3. Example

Scenario A — Privileged Session Without Recording Prevents Incident Investigation: An AI agent with production write access to the general ledger executes a series of journal entries during a month-end close process. Three weeks later, an internal audit identifies a £4.2 million variance that traces to journal entries posted during that session. The organisation attempts to investigate but finds that the agent's session was not recorded beyond basic action logs. The action logs show that 847 journal entries were posted but do not capture the agent's reasoning, the input data it processed, the intermediate calculations it performed, or the specific source documents it referenced. The investigation cannot determine whether the entries were correct but misclassified, incorrect due to a reasoning error, or the result of manipulated input data. The external auditor qualifies the financial statements.

What went wrong: The privileged session (production write to the general ledger) was not recorded with sufficient fidelity. Basic action logs captured the outputs but not the inputs, reasoning, or context. The organisation could not reconstruct what happened during the session, making investigation impossible and audit qualification unavoidable. Consequence: Qualified financial statements, regulatory scrutiny, £1.8 million in additional audit and investigation costs, CFO personal liability concern under SOX officer certification.

Scenario B — Tampered Session Recording Undermines Regulatory Response: A financial AI agent is investigated by the FCA for potentially manipulative trading patterns. The firm produces the agent's session recordings, but the recordings were stored in a database that the agent's operating team had write access to. The FCA's forensic analysis identifies gaps in the recording timeline and metadata inconsistencies suggesting that records were modified after the fact. The FCA treats the tampered recordings as an obstruction of its investigation and imposes a separate penalty for record-keeping failures in addition to the underlying trading investigation.

What went wrong: Session recordings were stored in a system modifiable by the operational team responsible for the agent. The recordings were not tamper-evident. When regulatory scrutiny arrived, the integrity of the records could not be demonstrated. Consequence: Separate FCA penalty for record-keeping failures (£2.7 million), enhanced scrutiny of all firm AI operations, personal liability proceedings against the head of AI operations under the Senior Managers Regime.

Scenario C — Session Recording Enables Rapid Root Cause Analysis: An AI agent with break-glass emergency access modifies production firewall rules during a security incident. The session is fully recorded: every command issued, every rule modified, the agent's reasoning for each modification, the input data it processed (security alerts, system state), and the timestamps of each action. Post-incident, the security team reviews the recording and identifies that the agent correctly mitigated the primary threat but inadvertently opened a secondary vulnerability by removing a rule that also protected an adjacent system. The team closes the secondary vulnerability within 2 hours of the incident resolution. Without the recording, the secondary vulnerability would have persisted undetected.

What went right: The privileged session was recorded with full fidelity, enabling rapid post-incident analysis. The recording captured not just the actions but the reasoning, enabling the team to understand why the agent made each decision and identify the unintended consequence.

4. Requirement Statement

Scope: This dimension applies to any AI agent session where the agent operates with elevated privileges relative to its routine operational permissions. Elevated privileges include but are not limited to: production write access (AG-302), administrative operations on infrastructure or systems, access to sensitive data classifications (e.g., Confidential, Restricted, classified), cross-workspace access (AG-299), break-glass emergency access (AG-306), access to PII or PHI beyond routine scope, financial transaction authority above routine thresholds, and any action that the organisation's risk classification designates as high-risk. Read-only operations within the agent's routine scope may be excluded from session recording, though organisations should define clear boundaries between routine and elevated operations. The recording obligation attaches to the session, not the individual action — the entire session in which elevated privileges are active is recorded.

4.1. A conforming system MUST record all agent sessions involving elevated privileges with sufficient fidelity to reconstruct the complete sequence of actions, inputs, outputs, reasoning (where available), and system state changes during the session.

4.2. A conforming system MUST store session recordings in a tamper-evident format that detects any post-capture modification, using cryptographic integrity protection (e.g., hash chains, digital signatures, write-once storage).

4.3. A conforming system MUST ensure that the recorded agent and its operating team cannot modify or delete session recordings — recordings are stored in a separate security domain with access restricted to audit, compliance, and security functions.

4.4. A conforming system MUST retain session recordings for the period required by applicable regulations and the organisation's data retention policy, with a minimum of 12 months online and 5 years archived.

4.5. A conforming system MUST include in each session recording: session identifier, agent identifier, privilege level, session start and end timestamps, every action executed (with parameters), every response received, input data processed (or a reference to it), and the outcome of each action (success, failure, or rejection).

4.6. A conforming system SHOULD capture the agent's reasoning chain or decision trace where the agent architecture makes this available (e.g., chain-of-thought outputs, tool selection rationale, intermediate reasoning steps).

4.7. A conforming system SHOULD implement real-time monitoring of privileged sessions, with alerts for anomalous actions during the session rather than exclusively post-session review.

4.8. A conforming system SHOULD protect sensitive data within session recordings through field-level encryption or redaction, ensuring that the recording itself does not become a data leakage vector.

4.9. A conforming system MAY implement session recording replay capability that allows auditors to step through the session chronologically, viewing the agent's inputs, actions, and reasoning at each step.

5. Rationale

Privileged sessions are where the highest-impact agent actions occur. Production writes, administrative changes, break-glass access, and sensitive data operations carry the greatest risk of harm if executed incorrectly and the greatest regulatory scrutiny when incidents occur. Session recording provides three critical capabilities: forensic investigation (understanding what happened after an incident), real-time oversight (monitoring high-risk operations as they occur), and regulatory evidence (demonstrating compliance with record-keeping obligations).

The recording obligation is proportionate to the privilege level. Routine read operations within the agent's normal scope carry low risk and generate high volume — recording everything would be prohibitively expensive and would dilute the signal. Elevated privilege sessions carry high risk and occur less frequently — recording them at high fidelity is both feasible and essential.

Tamper-evidence is non-negotiable for regulatory purposes. A session recording that could have been modified after the fact has no evidentiary value. Regulators treat modifiable records as unreliable, and the absence of tamper-evident records can itself constitute a regulatory violation. The recording must demonstrate, through cryptographic means, that the record as produced for review is identical to the record as captured during the session.

The separation of recording storage from the operational team is equally critical. If the team that operates the agent can modify the session recordings, the recordings cannot be trusted for audit or investigation purposes. The recording system must be architecturally independent — separate infrastructure, separate access controls, separate administrative authority.

6. Implementation Guidance

Session recording for AI agents extends traditional privileged access management (PAM) recording to the AI-specific context: capturing not just actions but reasoning, context, and decision traces.

Recommended patterns:

Anti-patterns to avoid:

Industry Considerations

Financial Services. MiFID II Article 16(7) requires firms to record telephone conversations and electronic communications relating to transactions. For AI trading agents, session recording extends this obligation to the agent's decision process. The FCA expects that firms can reconstruct the rationale for every trade, including trades executed by AI agents. Session recordings must be retained for 5 years (7 years at FCA request).

Healthcare. HIPAA audit controls (§164.312(b)) require mechanisms to record and examine activity in information systems containing ePHI. Session recording for AI agents accessing ePHI implements this requirement. Recordings containing ePHI must themselves be protected as ePHI.

Government and Defence. Audit requirements for classified system access typically require keystroke-level recording. AI agent session recording should capture the equivalent fidelity: every input processed, every action taken, and every output produced. Retention periods for classified systems may be indefinite.

Maturity Model

Basic Implementation — Privileged agent sessions are logged with action-level detail: action type, timestamp, target resource, success/failure. Logs are stored in a centralised logging system with access controls separating operational and audit access. Logs are retained for 12 months online. Limitations: no reasoning or context capture; log storage is mutable (not tamper-evident); no real-time monitoring of privileged sessions.

Intermediate Implementation — Privileged sessions are recorded with full fidelity: actions, inputs, outputs, and reasoning traces (where available). Recordings are stored in write-once storage with hash chain integrity. The recording system is architecturally separate from the operational team's infrastructure. Real-time monitoring generates alerts for anomalous actions during privileged sessions. Session recordings are searchable by agent, time period, action type, and privilege level. Field-level encryption protects sensitive data within recordings.

Advanced Implementation — All intermediate capabilities plus: session recording has been verified through independent adversarial testing including tamper attempts, recording bypass attempts, and evidence reconstruction challenges. Session replay capability enables auditors to step through sessions chronologically. Automated anomaly detection analyses privileged sessions against behavioural baselines and flags deviations. External timestamping services anchor recording integrity. The organisation can demonstrate to regulators that any privileged agent action can be fully reconstructed from the recording.

7. Evidence Requirements

Required artefacts:

Retention requirements:

Access requirements:

8. Test Specification

Test 8.1: Recording Completeness for Privileged Sessions

Test 8.2: Tamper Evidence Verification

Test 8.3: Recording Independence From Operational Team

Test 8.4: Recording Bypass Prevention

Test 8.5: Sensitive Data Protection in Recordings

Test 8.6: Retention Period Enforcement

Test 8.7: Real-Time Monitoring Alert Generation

Conformance Scoring

9. Regulatory Mapping

RegulationProvisionRelationship Type
MiFID IIArticle 16(7) (Record-Keeping of Transactions)Direct requirement
FCA SYSC9.1.1R (Record-Keeping)Direct requirement
SOXSection 802 (Criminal Penalties for Altering Documents)Supports compliance
HIPAA§164.312(b) (Audit Controls)Direct requirement
NIST SP 800-53AU-3 (Content of Audit Records), AU-9 (Protection of Audit Information)Direct requirement
EU AI ActArticle 12 (Record-Keeping)Direct requirement
GDPRArticle 30 (Records of Processing Activities)Supports compliance
DORAArticle 10 (Detection)Supports compliance

MiFID II — Article 16(7) (Record-Keeping of Transactions)

Article 16(7) requires investment firms to arrange for records to be kept of all services, activities, and transactions sufficient to enable the competent authority to fulfil its supervisory tasks. For AI trading agents, this extends to the agent's decision process — not just the trades executed but the reasoning, inputs, and decision logic that led to each trade. Session recordings of AI agent trading sessions provide the evidence base for regulatory reconstruction of trading decisions. Recordings must be retained for 5 years, extensible to 7 years at the FCA's request.

EU AI Act — Article 12 (Record-Keeping)

Article 12 requires providers of high-risk AI systems to ensure that the AI system is designed and developed with capabilities enabling automatic recording of events (logs). The logs must be sufficient to enable monitoring of the AI system's operation and to facilitate post-market monitoring. Privileged session recording implements Article 12 for the most consequential agent operations, providing the log fidelity needed for both operational monitoring and post-incident investigation.

SOX — Section 802 (Criminal Penalties for Altering Documents)

Section 802 imposes criminal penalties for knowingly altering, destroying, or concealing records with the intent to impede a federal investigation. Tamper-evident session recordings protect the organisation against Section 802 liability by ensuring that recordings cannot be altered — the integrity protection makes any alteration detectable, and write-once storage prevents deletion. This provides both a compliance safeguard and a legal defence.

NIST SP 800-53 — AU-3, AU-9

AU-3 requires audit records to contain sufficient content to establish the type of event, time, source, outcome, and identity. AU-9 requires protection of audit information from unauthorised access, modification, and deletion. Session recordings with full fidelity satisfy AU-3, and tamper-evident storage with access separation satisfies AU-9.

10. Failure Severity

FieldValue
Severity RatingHigh
Blast RadiusAccountability — loss of ability to investigate incidents, satisfy regulatory inquiries, or demonstrate compliance for privileged operations

Consequence chain: Failure of privileged session recording does not directly cause an operational incident — it prevents the organisation from understanding, investigating, and responding to incidents that occur during privileged sessions. The consequence materialises when an incident occurs and the organisation cannot reconstruct what happened. For financial services, inability to reconstruct an AI trading agent's decisions may constitute a MiFID II violation (£millions in fines). For SOX-regulated firms, inability to explain AI-generated financial entries may lead to audit qualification and officer liability. For healthcare, inability to explain AI actions on patient records may constitute a HIPAA audit control violation. Beyond regulatory consequences, the absence of recordings prevents organisational learning — the organisation cannot improve its governance controls if it cannot understand how they failed. The blast radius extends to organisational trust: if the organisation cannot demonstrate what its AI agents did, stakeholders (regulators, clients, partners, employees) cannot trust the AI programme.

Cross-references: AG-302 (Production Write Isolation Governance) provides the preventive controls that session recording complements for production writes. AG-304 (Just-in-Time Secrets Issuance Governance) defines the privilege issuance events that trigger recording obligations. AG-306 (Break-Glass Containment Governance) defines emergency access sessions that require the highest recording fidelity. AG-013 (Data Sensitivity and Exfiltration Prevention) ensures that session recordings themselves do not become data leakage vectors. AG-162 (Least-Agency Provisioning) minimises the frequency of elevated privilege sessions, reducing recording volume.

Cite this protocol
AgentGoverning. (2026). AG-305: Privileged Session Recording Governance. The 783 Protocols of AI Agent Governance, AGS v2.1. agentgoverning.com/protocols/AG-305