GitHubSign inStart testing

Complete trace analysis

See what your agent exposes on the way to the answer.

AgentLeak replays the whole execution trace and follows every sensitive value through 8 normalized channels. A clean final answer no longer hides a leak in a tool call, shared memory, a log or a generated file.

Create a workspace Agents: discover and onboard

8 channels, one schema

Traces from any framework normalize to one AgentLeak schema, so the analysis is identical whatever produced the run.

  • User input and final output
  • Tool calls and tool responses
  • Inter-agent messages and shared memory
  • Logs and generated files

Severity you can defend

Every finding gets a level from L1 to L4 based on how sensitive the exposed value is and where it went, not a vague red/yellow/green badge.

  • L1 to L4 severity per finding
  • The exact channel where exposure happened
  • Reconstructed leak path across events
  • Canary matching for realistic vaults

A fix, not just a flag

Each finding carries a remediation hint: prose for your team and a structured, machine-readable hint an agent can apply on its own.

  • Ready-to-paste code fixes
  • Per-channel redaction advice
  • Priority-sorted next steps
  • Structured hints for autonomous agents

Inside the capability

Three views of how it works.

NORMALIZED INPUT
INPUTTOOLSMEMORYHANDOFFLOGSOUTPUTRUN · 01HF7A18 EVENTS

Eight channels, one event clock

Align tool calls, memory, logs and outputs on the same ordered trace instead of reviewing separate exports.

VALUE PROVENANCE
SOURCEvault.emailAGENT BmemorySINK · L3tool_call10:42.0110:42.1810:42.31

Follow one value across boundaries

See where sensitive data entered, which agent handled it and the exact event where it escaped.

AUDITABLE EVIDENCE
FINDINGSLEVELtool_callcustomer.emailL4shared_memorypatient.idL3logssession.tokenL4final_output—PASSdigest · a91c…e204

Every finding keeps its proof

Channel, timestamp, level and redacted evidence stay attached to the finding for review and replay.

From execution to evidence

Turn every agent signal into a privacy action.

Auto-create findings from agent runs

Capture disclosures from tool calls, memory and handoffs, then create trace-linked remediation without prompts or manual review.

Create privacy tests from every execution surface

Ingest frameworks, OpenTelemetry, MCP, logs and CI, then normalize every source into one auditable AgentLeak run.

The mental model

An agent run is a dataflow graph, not a chat transcript.

A final answer is only one exit from the system. Sensitive data can enter through a tool response, move through memory, cross an agent hand-off and reach a third party without ever appearing in the answer. AgentLeak treats every event as an edge in that flow and reconstructs the complete disclosure path.

Before

Output checks ask: did the answer contain a secret?

With AgentLeak

Trace analysis asks: where did each sensitive value travel?

  • Sources establish what the agent was allowed to see
  • Disclosure channels establish what the agent emitted
  • Distinct values are followed across events, not counted as isolated strings

How it works

From raw trace to a fix, in four steps.

1

Capture the trace

Record events at trust boundaries: user input, tool calls and responses, memory, logs and the final output. Any framework works.

2

Normalize to 8 channels

LangChain, LangGraph, CrewAI, MCP, OpenTelemetry and generic OpenAI-style logs all map to the same AgentLeak schema before analysis.

3

Replay and match

Each channel is scanned with regex, Presidio, entropy and de-obfuscation detectors, matched against the vault, and reconstructed into a leak path.

4

Read the report

Every finding lists the exact channel, a severity from L1 to L4, and a fix, in the CLI, the dashboard or a JSON report.

Analyze a trace from the CLI

agentleak run --trace run.json --output ./reports --format html
open ./reports/run_0001.html

Read the complete implementation guide →

Explore the platform

The rest of the loop.

Static code scanCatch secrets before the agent runsLearn more MCP preflightAgents check what changed before they finishLearn more Adversarial red-teamReplay real attack classes, not just clean tracesLearn more

FAQ

Questions, answered.

Ready when you are

Test the path, not only the answer.

Create a local workspace, run a bundled scenario, then wire AgentLeak into CI or let your agent onboard itself.

Create a workspace Agents: discover and onboard