Ready reckoner
Readiness
- Lifecycle unchanged: prepare → detect/analyze → contain → eradicate → recover → learn. Contents change.
- RACI across SOC, MLOps, model owner, data owner, privacy, legal, comms.
- Release requirements include two-sided baselines (override rate too high or too low is a signal).
- Prompts, guardrail rules and tool definitions as configuration as code, pinned, with drift detection.
Telemetry
| Layer | Must log |
|---|---|
| Gateway | Identity, prompt (masked/hashed), tokens, filter verdicts |
| Agent | Plan, steps, tool chosen, approvals, on-behalf-of identity |
| Retrieval | Query, chunk IDs/sources, permission-check result |
| Model | Exact version, settings, prompt version, output |
| Tools | Tool, parameters, result, credential used |
- Join by trace ID; sync clocks; send to SIEM; SOAR for reversible first responses.
- Logs as evidence: separate store, immutable, hashed, gap monitoring, chain of custody.
- Privacy tiers: metadata (broad) · masked content with consistent tokenization (analysts) · full content (break-glass).
Detection
- Points: inputs, retrieval, outputs, tools, resource use. Layer signatures + classifiers + anomaly detection; tune on misses and false alarms.
- Canary tokens in prompts/documents → extraction detected.
- Access misuse (wrong identity) → revoke credential. Agency misuse (right agent, wrong action, often after untrusted content) → pause agent, quarantine content.
- Exposed secret → revoke and rotate first, then trace spread.
- UEBA with per-agent baselines; SOC AI protected from injection via logs it reads.
Vulnerability posture
- Registers: vulnerability + exposure. Many AI weaknesses have no CVE, no patch.
- Verified closed = many attempts and variants vs a criterion set in advance.
- Prioritize by exposure, exploitation, data, privileges, existing controls, consequence; CVSS + KEV + EPSS for conventional flaws.
- Legacy models: compensating controls, usage restrictions, time-limited exceptions with expiry.
Response
- Taxonomy: injection, jailbreak, poisoning, extraction, leakage, agent misuse, supply chain, harmful output.
- Involve legal and privacy early: CERT-In 6 h · NIS2 24 h · GDPR 72 h · AI Act serious incidents.
- Containment ladder: tighten guardrails → throttle/block → reduce capability → isolate → roll back → kill switch. Least disruptive step that reliably stops harm.
- Stop agents: revoke credentials → hold pending actions → record state → stop.
- Evidence in order of volatility; snapshot before cleaning.
Investigation and eradication
- Kill chains: poisoning (trace with lineage) · indirect injection (trace with traces). ATLAS vocabulary.
- Trace replay from logs, not re-runs. Artifact comparison, dataset diffing, isolated behavioral tests, unified timeline.
- Eradicate data, models, derived models (fine-tunes, distillations, adapters), indexes, caches, agent memory, enabling weakness, attacker access.
- Reintroduce via regression tests, red-team retest, verification, staged rollout.
Recovery
- Verified trusted state: signed rollback, clean checkpoints, rebuild index (not piecemeal deletes), configs from version control.
- Circuit breakers and kill switches act outside the model, fail safe, are tested.
- BCP: dependency map, RTO/RPO, tested alternatives, degraded modes, tested restores.
- Blameless reviews → tracked actions, TTP register, detections, playbooks, tests.
- Data removal: unlearning (fast, approximate) · retraining (certain, costly) · termination.