Aileron

Aileron is a flight recorder for AI agents.

Not another tracer. Aileron produces a tamper-evident, replayable record of every tool call your agents make - evidence you can verify offline, not telemetry you have to trust.

• Tamper-evident audit trail. Every agent action is appended to a SHA-256 hash-chained JSONL journal with Ed25519-signed checkpoints. Edit, delete, or reorder a single line and aileron verify says exactly where the chain broke.
• Policy enforcement on tool calls. Sigma-like YAML rules with allow / alert / block actions, applied before execution via the MCP stdio proxy or the SDK decorator. A blocked tool call never runs; the attempt is logged anyway.
• Forensic incident replay. One command turns a journal into a self-contained HTML incident report with a verification badge and a filterable timeline - the answer to "what did the agent actually touch?"

60-second quickstart

$ pip install aileron
$ aileron demo            # scripted fake-agent session (no network, no keys needed)
demo: wrote 8 events to demo.chain.jsonl
demo: chain VERIFIED (8 events)
demo: blocked shell call by rule aileron-001
demo: 2 anomaly alert(s) emitted
$ aileron verify demo.chain.jsonl
OK: 8 events verified in demo.chain.jsonl
$ aileron report demo.chain.jsonl -o incident.html   # open it in a browser
$ aileron serve --root .                             # or ask an assistant instead

The demo runs in the default digest-only mode: the destructive shell call is blocked by a content rule and flagged by the behavioral baseline, yet the journal on disk contains only argument digests - never the raw command.

Features

Usage

SDK: @track decorator

from aileron import ChainLog, track, PolicyBlocked, bundled_rules_dir
from aileron.policy import load_rules

log = ChainLog("run.chain.jsonl")            # capture_content=False by default
rules = load_rules(bundled_rules_dir())      # or load_rules("rules") after `aileron init`

@track(log=log, rules=rules)
def shell(cmd: str) -> str:
    ...  # your tool implementation

shell("ls /tmp")            # -> tool_call event, status=ok, args recorded as digest
shell("rm -rf /")           # -> PolicyBlocked raised; blocked attempt is logged

Rules see the full arguments in memory at decision time; the journal still stores digests only. Turn on capture_content=True only when you want raw arguments persisted for forensics.

SDK: track_agent session

from aileron import track_agent

with track_agent("research-agent", framework="langchain", log=log):
    shell("ls /tmp")   # inherits the session's agent identity and session_id
# agent_start / agent_end events bracket the run automatically

MCP proxy: framework-agnostic interception

Wrap any MCP server. Every tools/call is logged and policy-checked before the child process sees it:

$ aileron init                       # seeds a ./rules directory with starter rules
$ aileron proxy --log run.chain.jsonl --rules rules -- \
    npx -y @modelcontextprotocol/server-filesystem /tmp

A blocked call returns a JSON-RPC error (-32000: blocked by aileron rule <id>) to the client; the child is never invoked.

Verified against real MCP servers, not just test doubles. Aileron has been run in front of the official @modelcontextprotocol/server-filesystem (secure-filesystem-server 0.2.0, 14 tools) and @modelcontextprotocol/server-memory (0.6.3, 9 tools): the handshake completes, tools list normally, real calls work, a blocked write never reaches the server, and the journal verifies. That check ships as a test (tests/test_real_mcp_server.py, run with AILERON_LIVE_MCP=1).

The proxy itself costs well under a millisecond per tools/call. Matching content rules against large payloads costs more, and how much is yours to choose: see Performance for the split, measured.

The proxy speaks both newline-delimited and Content-Length-framed JSON-RPC. Content rules (tool.arguments_contains, _regex) work in the default digest-only mode - --capture-content changes what is persisted, not what is enforced. Calls still in flight when the child dies are journaled with status=error, so a crash never erases the attempt.

MCP server: ask your assistant what the agent did

Aileron sits in front of MCP servers. It is also one. Point it at a directory of journals and an assistant can read the record for you:

$ aileron serve --root ./journals

Three tools, all read only: verify_journal (is this record intact), query_events (what happened, filtered by tool, status, or time), and explain_rule (what does aileron-130 catch).

There is no write, delete, or sign tool, and there should never be. The agent being recorded is the untrusted party, so giving it a way to edit the journal would hand the suspect the evidence locker.

Four things follow from that, and they are the reason this is more than a wrapper around aileron verify:

• Paths are confined to --root and only .jsonl opens. Otherwise verify_journal(path) is an arbitrary file read.
• Every answer carries its own integrity status. Confinement stops an agent reading files it should not; it does not stop one writing a plausible journal inside the root and handing you an invented history. So each reply says whether the chain verifies and whether an adjacent signed checkpoint agrees.
• Recorded values are treated as hostile. Tool names and rule ids were chosen by the agent under investigation, so they reach an assistant labelled as untrusted data, stripped of control characters, and truncated. A tool named IGNORE PREVIOUS INSTRUCTIONS... is evidence to report, not an instruction to follow.
• Digest-only stays digest-only. capture_content governs what the journal stores. It never widens what this server hands back, and errors never echo file contents.

Policy rules

# a policy rule (see the bundled rules/examples/destructive-shell.yml)
id: aileron-001
title: Block destructive shell commands
severity: high
match:
  type: tool_call
  tool.name: shell
  tool.arguments_contains: ["rm -rf", "DROP TABLE", ":(){ :|:& };:"]
action: block

Dry-run rules against a recorded session: aileron rules test rules/ run.chain.jsonl

How it works

agent ──tool call──► [ SDK @track ] ──┐
                     [ MCP proxy  ] ──┼─► policy decide (allow/alert/block)
                                      │        │ block? ──► call never executes,
MCP client ──JSON-RPC──► proxy ───────┘        │        attempt still logged
                                               ▼
                              append to chain log (JSONL)

  event 0           event 1                      event N
 ┌──────────────┐  ┌──────────────┐        ┌──────────────┐
 │ seq: 0       │  │ seq: 1       │        │ seq: N       │
 │ prev: 0000…  │─►│ prev: H(e0)  │─► … ──►│ prev: H(eN-1)│
 │ hash: H(e0)  │  │ hash: H(e1)  │        │ hash: H(eN)  │──► Ed25519 checkpoint
 └──────────────┘  └──────────────┘        └──────────────┘    signature over tip

  H(e) = sha256(canonical_json(e \ hash))
  aileron verify          → recompute every hash + link (exit 2 on tamper)
  aileron verify-checkpoint → re-verify chain tip against Ed25519 signature

Tampering with any event breaks the hash link at the first modified sequence; verify reports first_bad_seq and exits non-zero. The journal is local-only and self-contained - verification needs no network and no trusted third party.

Performance

The proxy adds well under a millisecond per tools/call. Matching the full 32-rule bundled pack against the payload is a separate cost that grows with payload size, and it is reported separately below, because the two scale differently and you choose your own rules.

Every number is reproducible with one command:

$ python scripts/benchmark.py

Method. scripts/benchmark.py drives an identical stdio MCP child server three ways - directly, through aileron proxy with no rules, and through aileron proxy with all 32 bundled rules - and subtracts. The deltas are the proxy's true cost, so you never have to trust an absolute figure. The absolute baseline is printed alongside so the subtraction can be checked. 2,000 sequential calls per configuration after 200 discarded warmup calls; digest-only journaling. Overhead covers JSON-RPC parsing, policy evaluation, hash-chain append, re-serialization, and the extra process hop.

About the payload. The tool arguments are fixed text that looks like real tool arguments: English words, paths, flags, quotes and punctuation. That matters more than it sounds. This benchmark used to send a run of one repeated character, which is the friendliest possible input both to the regex engine, which fails on the first character everywhere, and to the literal prefilter described below, which finds nothing anywhere. It was flattering the result by about 3x. A test asserts no bundled rule fires on the filler, so these numbers are the ordinary path and not the alert path.

Added latency per tools/call (milliseconds)

Linux x86_64 - GitHub Actions ubuntu-latest (2 shared vCPU), Python 3.12.14. Re-measured by CI on every push:

Shared CI runners vary between runs, by as much as 1.6x on the small-payload row. This table quotes the slower of two consecutive measurements. The regression baseline uses the faster one, so a slow runner cannot quietly widen the guard.

macOS arm64 - Apple M2 Pro, Python 3.13.7, idle machine. The worst of three passes, quoted whole, so added = (proxy & rules) - direct holds exactly within the run:

What these numbers mean

The proxy is cheap and nearly flat. Interception, journaling, and re-serialization cost about 0.23 ms on a small call and about 0.56 ms on a 32 KB one, on the slowest hardware tested.

Rules cost more on big payloads, and the cost is yours to choose. Content rules are matched against the payload, so their cost grows with payload size. With all 32 bundled rules loaded that is 0.24 ms on a small call and 2.23 ms at 32 KB.

Most of that work is skipped before it starts. A rule looking for auditctl cannot fire on a payload with no auditctl in it. Each pattern is read once and reduced to the literals it requires, and cheap substring searches decide whether the regex runs at all. Requirements are conjunctions, so a rule needing systemctl near disable near auditd is skipped on ordinary prose that merely contains the word "service". On a benign 32 KB call, 17 of the 19 patterns that would otherwise scan the whole payload never run. It changes speed and nothing else, and AILERON_NO_PREFILTER=1 turns it off if you want it ruled out during an investigation.

In context. A real MCP server call is typically 10 to 1000 ms. At 2.8 ms for a 32 KB argument with every rule loaded, and 0.47 ms for an ordinary small one, mediation is a small fraction of the call it is mediating.

Caveats, stated plainly. These are sequential stdio round-trips, one call in flight at a time, which is how an agent actually calls tools. This is not a concurrent-client benchmark; a many-client run is on the roadmap. The tool reports mean, median, p95 and p99; these tables quote medians, because medians are stable across runs and p95 is not - tail latency swings with scheduling. Linux is the slower machine because a shared-vCPU CI runner is slower than an idle laptop, and those are the conservative figures CI enforces. Measure on your own hardware before quoting a number.

CI enforces this: a job fails if median overhead regresses more than 2x against scripts/benchmark_baseline.json, so performance cannot decay silently. It re-measures once before failing, so a single slow runner does not cry wolf. The baseline records both the rule-pack size and the payload shape it was measured against, because a change to either is more work rather than slower code, and the guard says so instead of reporting a regression that is not there.

Integrations & ecosystem

• OpenTelemetry GenAI - aileron export emits gen_ai.operation.name / gen_ai.tool.name / gen_ai.agent.name span attributes plus aileron.event.hash, so Aileron sits beside your existing tracing stack as the evidence layer, not instead of it.
• LangChain / CrewAI / any Python framework - @track is a plain decorator; track_agent accepts a free-form framework= label. No framework dependency is required.
• MCP - the proxy wraps any stdio MCP server regardless of which client or framework drives it.
• Community rules - rule contributions are the intended contribution unit (see Roadmap).

Telemetry & privacy

• Aileron sends no telemetry. No analytics, no phone-home, no network calls anywhere in the library or CLI. If that ever changes, it will be opt-in only, behind a documented RFC - for a security tool, anything less is disqualifying.
• Content capture is off by default. Tool arguments and results are recorded as SHA-256 digests; raw content is only stored when you pass capture_content=True / --capture-content. You get a verifiable record of what happened without persisting secrets or PII by accident. Policy rules and the anomaly detector still see the full call in memory at decision time - capture only controls what is persisted, never what is enforced.

Honest limitations

• SDK instrumentation is bypassable. @track wraps the functions you decorate; code paths you don't instrument are not recorded. For enforcement that agent code cannot skip, use the MCP proxy - mediation happens in a separate process on the tool-call path.
• Policy rules are pattern matching, not intent classification. They catch known-bad shapes (rm -rf, id_rsa, exfil patterns); they will not reliably detect novel malicious reasoning. Detection-of-effect complements detection-of-intent tools (garak, PromptGuard); it does not replace them.
• Tamper-evidence is not tamper-proof. The chain proves modification after the fact; an attacker with write access can truncate or rewrite the whole log and forge it forward. Signed checkpoints make forgery require the private key - keep keys off the host being recorded, and anchor checkpoints externally (see Roadmap) if you need non-repudiation.

Roadmap

• aileron-rules community rule repo - Sigma-for-agents: community detection rules mapped to the OWASP Agentic Security Initiative's threat taxonomy, CI-validated against recorded incident traces.
• Sigstore/Rekor checkpoint anchoring - publish signed checkpoints to a public transparency log for non-repudiable, third-party-verifiable timestamps.
• OCSF export - emit Open Cybersecurity Schema Framework events for direct SIEM ingestion (Splunk/Elastic quickstarts).
• Out of scope for v1: eBPF / kernel-level interception. Aileron stays at the MCP-proxy and SDK layer where the semantic meaning of a tool call is still visible; syscall-level tracing is Falco/Cilium territory and would trade agent semantics for volume.

Contributing

Contributions are welcome - see CONTRIBUTING.md. Good starting points: new detection rules under src/aileron/rules/examples/ and new framework adapters under examples/. DCO sign-off, no CLA. Security issues: see SECURITY.md.

License

Apache License 2.0 - see LICENSE.