AgentTrace

AI agent forensics & incident response — reconstruct, verify, and report on LLM/agent security incidents.

Local-first · standards-aligned · zero runtime dependencies · defensive DFIR.

AgentTrace is an open-source digital forensics and incident response (DFIR) tool for AI agents and LLM applications. It ingests already-recorded evidence (OpenTelemetry GenAI spans, MCP server logs, vector-store retrieval logs, OAuth records, egress logs, Halo-record hash-chains), verifies its integrity and chain of custody, reconstructs the causal attack chain, detects known attack patterns mapped to MITRE ATLAS, and produces a regulator/court-ready report aligned with EU AI Act Article 12.

Keywords: AI agent forensics · LLM incident response · prompt-injection investigation · agentic AI security · MITRE ATLAS · chain of custody · OpenTelemetry GenAI · MCP · EU AI Act · DFIR.

What it does

• 🔌 Ingests 7 evidence sources into one normalized schema (OTel GenAI, MCP, vector store, OAuth, egress, Halo-record, generic JSONL).
• 🔒 Verifies integrity & chain of custody — hash chains, gap/witness detection, HMAC-signed manifests, a tamper-evident custody ledger.
• 🕸️ Reconstructs the causal attack chain across fragmented logs (timeline + provenance graph).
• 🎯 Detects 6 attack patterns (prompt injection via retrieval, exfiltration via tool chaining, OAuth/credential theft, sub-agent hijack, memory poisoning, tool-permission escalation) — each mapped to MITRE ATLAS and linked to exact evidence.
• 🧭 Builds kill-chain narratives + risk scores.
• 📄 Reports in JSON, Markdown, and self-contained offline HTML (with an SVG causal-graph), plus EU AI Act Article 12 coverage.
• 📦 Portable signed case bundles for air-gapped transfer.
• 🚫 No AI/ML in the detection path — deterministic and fully explainable.

It is the investigation layer that sits downstream of recorders (OpenTelemetry GenAI, Halo-record, MCP logs). Recording is largely solved; automated reconstruction of what actually happened across 6–7 fragmented log sources is not. That is the gap AgentTrace fills.

⚠️ Authorized use only

AgentTrace is a defensive forensic tool. Use it only on evidence you are legally authorized to investigate (your own agent deployments, or incidents you are properly engaged to investigate). It is not an attack, exploitation, or surveillance tool, has no capability to access remote systems or accounts, and runs fully offline.

Why it exists

• The attack vector in AI-agent incidents is natural language — a malicious prompt looks identical to a legitimate one in every log. The sequence is the evidence, not any single event.
• Evidence is scattered across LLM invocation logs, tool traces, MCP server logs, vector-store retrieval logs, OAuth records, and egress logs.
• Volume defeats manual review (a documented 2026 incident: ~17,600 agent actions in 5 days; the operator called manual reconstruction "impractical").

Pipeline

ingest → normalize (UFE) → verify (integrity + chain of custody)
       → correlate (timeline + causal graph) → detect (attack patterns)
       → report (JSON + Markdown, EU AI Act Art.12 coverage)

• Unified Forensic Event (UFE): every source is normalized to one schema.
• Integrity: verifies Halo-record hash chains, detects sequence/time gaps, and supports out-of-band witness anchors — distinguishing "nothing was edited" from "nothing is missing".
• Correlation: builds a causal graph with PROV-O-style edges (parent_of, followed_by, used_data, derived_from).
• Detection (explainable, evidence-linked):
  1. prompt injection via retrieved content,
  2. exfiltration via tool-call chaining (monotonic-offset paging),
  3. OAuth / credential-theft chains with anomalous operational velocity.
• Reporting: chain-of-custody attestation + EU AI Act Article 12 coverage.

Install

pip install -e .          # from the project root
# or run without installing:
python3 -m agenttrace.cli --help

Requires Python ≥ 3.10. No third-party runtime dependencies.

Quickstart

Generate a synthetic incident dataset and run the full pipeline:

python3 -m tests.synthetic synthetic_data      # writes sample evidence files
agenttrace report synthetic_data/*.json synthetic_data/*.jsonl \
    --signing-key "my-case-key" \
    --title "Synthetic AI Agent Incident" \
    --md-out report.md --json-out report.json --html-out report.html

The --html-out report is fully self-contained and offline (inline SVG causal-graph visualization, no external CDN or network access).

Deterministic by design (no AI/ML in the detection path)

AgentTrace uses only deterministic heuristics and rule-based correlation — no LLM or ML model participates in reaching a finding. This is intentional for a forensic tool: every finding links back to the exact underlying events, there is no hallucination or automation-bias risk, and it runs air-gapped with zero dependencies. Any future AI assistance would be confined to an optional natural-language summary layer on top of the deterministic core, and would never decide a detection.

Other subcommands:

agenttrace ingest      <files...>   # normalize evidence to UFE (JSON)
agenttrace verify      <files...>   # integrity + chain of custody + manifest
agenttrace reconstruct <files...>   # timeline + causal graph
agenttrace detect      <files...>   # attack-pattern findings

Force a specific collector with --collector otel_genai|halo_record|jsonl_llm.

Supported evidence sources

Collector	Source format
otel_genai	OpenTelemetry GenAI spans (gen_ai.*)
halo_record	Halo-record hash-chained audit logs (JSONL)
mcp	Model Context Protocol server logs (JSONL)
vector_store	RAG vector-store retrieval logs (JSONL)
oauth	OAuth grant / token issuance & use (JSONL)
egress	Egress network logs (JSONL)
jsonl_llm	Generic JSONL LLM/tool logs (permissive)

New sources are added by subclassing Collector (see agenttrace/collectors/).

Tests

python3 -m unittest discover -s tests -v

All tests use synthetic data only; no real systems or data are involved.

Project layout

agenttrace/
  model.py         # UFE schema, evidence bundle, redaction
  custody.py       # tamper-evident chain-of-custody ledger
  collectors/      # pluggable ingestion (otel/halo/mcp/vector/oauth/egress/jsonl)
  integrity.py     # hash-chain verify, gap detection, signed manifest
  correlate.py     # timeline + causal graph
  detect.py        # attack-pattern detectors + MITRE ATLAS mapping
  analyze.py       # kill-chain narratives + risk scoring
  bundle.py        # portable signed .tar case bundles
  report.py        # JSON / Markdown / HTML reporting
  cli.py           # command-line interface
tests/             # synthetic dataset generator + test suite
DESIGN.md          # architecture & threat model

Contributing & security

• CONTRIBUTING.md — dev setup, adding collectors/detectors, style.
• SECURITY.md — responsible disclosure and safe evidence handling.
• CHANGELOG.md — release history.

Support

AgentTrace is free and open source. If you find it useful and would like to support development, donations are welcome — entirely at your own discretion.

• TRON (TRX / USDT–TRC20): TYqSCXX8Vu7MXcTXsPPgiKxb4uRHLKTPcC

⚠️ Send only on the TRON (TRC-20) network. Always verify the address character-for-character before sending. Cryptocurrency transactions are irreversible; donate at your own risk. Donations are voluntary and non-refundable.

License

Apache-2.0. See LICENSE.