Pipelock
Pipelock is an open-source (Apache 2.0 core) AI agent firewall that sits between agents and the network, mediating HTTP, WebSocket, MCP, and A2A traffic to detect secret exfiltration, prompt injection, SSRF, and MCP tool poisoning. It scans request/response bodies with a DLP engine (65 built-in patterns, checksum validation), streaming SSE inspection, and MCP tool-call chain analysis, then emits Ed25519-signed, hash-chained action receipts so decisions can be verified offline outside the agent runtime. It also offers OS-level containment (Landlock/seccomp/sandbox-exec), a kill switch, canary tokens, and integrations for Claude Code, Cursor, VS Code, OpenAI Codex, LangGraph, CrewAI, and other agent frameworks.
What it is
Pipelock v3.5.0 (Community edition, Go 1.25+), an open-source AI agent egress firewall. The summary covers its scanning scope; this page adds the concrete verification model, the build/licensing split, and several self-stated honesty caveats an evaluator needs before trusting its receipts.
How it works
Pipelock mediates outbound/inbound traffic and emits mediator-signed action receipts over boundary decisions, verifiable offline with a key you hold — no account, no server. It writes a "flight-recorder" session while running (distinct from the demo receipts). Detections are exercised against the public agent-egress-bench corpus; the "Gauntlet" workflow is a scheduled exam against a pinned corpus commit and does not auto-publish a public score.
Preconditions / Affected
- TLS interception must be enabled to inspect CONNECT tunnel contents. Plain CONNECT without interception is scanned only at hostname/URL level — a real coverage gap.
- Source builds (
make build/make install) produce a Community-only binary. Pre-built release artifacts contain paid-tier code that activates with a license key. The Pro/Enterprise dashboard, RBAC, and fleet views exist only in enterprise-tagged builds.
What to do
Zero-config demo that writes signed receipts and the public key, then verify a signature yourself:
pipelock demo --receipts-dir ./out
pipelock verify-receipt "$(ls ./out/*.json | head -1)" --key ./out/signer.pubRecord a live session and produce an offline report:
pipelock init --output ./pipelock.yaml
pipelock run --config ./pipelock.yaml
pipelock evidence view --receipt-dir ./recorder --out report.htmlVerify release artifact integrity (SLSA provenance, SBOMs, signed images published per release):
gh attestation verify pipelock_3.5.0_linux_amd64.tar.gz --owner luckyPipewrench
gh attestation verify oci://ghcr.io/luckypipewrench/pipelock:3.5.0 --owner luckyPipewrenchFree pipelock report, Prometheus, and Grafana paths monitor a single instance; cross-agent fleet control (Conductor) is Enterprise-only.
Caveats
The page states these limits itself:
- The demo signs with an ephemeral key it prints per run — receipts are self-consistent, not tied to a named identity. The public playground verifies against a key Pipelock publishes.
- The operator holds the signing key. A receipt proves what the boundary decided and that the key holder signed it — not that the operator is honest.
pipelock anchor receiptswrites receipt-chain checkpoints to a local backend or Rekor, but operator-independent verification against that anchor is "still being proven end to end."- Receipts prove nothing about activity outside the boundary Pipelock mediates.
Organisations named here
- OWASP Agentic AI Top 10
Technologies named here
- AI agents with API keys and shell access
- Pipelock
- Uncensored or abliterated LLM models
- Agent Egress Bench
- Gauntlet
- pipelock demo
- pipelock verify-receipt
- pipelock evidence view / serve
- pipelock init
- pipelock run
- pipelock anchor receipts
- pipelock check
- gh attestation verify
- pipelock dashboard serve
- pipelock report
- Prometheus and Grafana
- pipelock audit
- pipelock sandbox
- pipelock contain
- pipelock explain
- pipelock canary
- pipelock skill-scan
- pipelock mcp proxy
- pipelock posture emit / verify
- Conductor
- pipelock baseline
- Kill switch
- Scan API
- Filesystem sentinel
- pipelock assess
- pipelock rules
- Rekor
- docs/bypass-resistance.md
- SIEM/webhook/syslog/CEF/OTLP integrations
- Model Context Protocol
- Claude Code
- Cursor
- VS Code
- Docker
- Kubernetes
Techniques named here
- Secret exfiltration
- Prompt Injection
- SSRF
- Tool poisoning
- Risky tool-call chains
- CRLF injection / path traversal
- DNS rebinding
- Cryptocurrency address lookalike swap
- Credential-to-network-sink / shell-to-write
- Destructive tool calls
- Guardrail-removal toolchain detection
- Cross-request exfiltration detection
- Filesystem secret exfiltration
Related in AI Security
