Quick start · How it works · Safety filter · Real-world test · Survey

---

Your AI app's monitoring data (OpenTelemetry traces and logs) records prompts, tool calls and errors, and sends them to a dashboard. Secrets ride along: an API key in an error message, a password a user pasted into the chat, a token in a URL. Redaction code passes its unit tests while the key still leaks through a field nobody thought about.

TraceLeakCheck plants a fake secret, captures everything your app really exports, and tells you in plain English exactly where it leaked.

What you get

🔎 Find	pytest --traceleak checks every test that uses the canary fixture: traces, logs and metrics, over HTTP or gRPC, raw or encoded (base64 at any offset, URL, hex, truncated).
🛠️ Fix	scrub() wraps your exporter and removes secrets before anything leaves your app. Every report comes with how-to-fix tips.
✅ Prove	Re-run and get clean, backed by a control check showing your telemetry really arrived. "Nothing arrived" is never "nothing leaked".

Quick start

pip install traceleakcheck

def test_agent_does_not_leak_key(canary):
    key = canary("openai").value          # a fresh fake secret; this test is now checked
    run_agent("summarize this", api_key=key)

pytest --traceleak                                 # plain-English results in the terminal
pytest --traceleak --traceleak-html=report.html    # + a report page for your browser
pytest --traceleak -v                              # + exact technical field paths

If your app configures OpenTelemetry from the standard OTEL_EXPORTER_OTLP_* settings, it needs no changes. Traceloop's TRACELOOP_BASE_URL is set too. If your app builds a provider by hand, register it with traceleak.watch(provider).

The HTML report

How it works

1. Before your conftest.py loads, --traceleak starts a receiver and points the app's exporters at it.
2. Each test that asks for a canary gets unique fake secrets in realistic formats: OpenAI, Anthropic, Azure OpenAI, AWS, GitHub, JWT or password.
3. After the test and its fixtures finish, a control marker is sent through every configured pipeline (traces, logs, metrics), everything is flushed, and every exported field is searched.

Each test gets one verdict:

Verdict	Meaning	Test
✅ clean	Every pipeline delivered its control marker, and no canary was found	passes
🚨 leak	A canary was found; the report names the step and field	fails
⚠️ incomplete	A pipeline's control marker never arrived, or a span was still open at the end of the test (open spans are never exported), so "clean" can't be claimed	fails
⚠️ error	An export couldn't be decoded, so a scan would under-report	fails

pytest shows a failed check as ERROR at teardown, because it runs after the fixtures finish. It works with pytest -n (xdist). In GitHub Actions the verdicts also go to the job summary, and there's a ready-made action.

Fix it: the safety filter

from traceleakcheck.guard import scrub

tracer_provider.add_span_processor(BatchSpanProcessor(scrub(OTLPSpanExporter())))
logger_provider.add_log_record_processor(BatchLogRecordProcessor(scrub(OTLPLogExporter())))

Before anything is exported, the filter removes:

• Known secrets: the values of environment variables named like *KEY*, *TOKEN*, *SECRET* or *PASSWORD*, plus anything registered with add_secret(). Their base64, URL and hex forms are removed too. Values that are plainly not secrets are left alone, so they don't vanish from every trace: file paths, URLs without a password, and, for *KEY*, *TOKEN*, *AUTH* and *CREDENTIAL* names, plain words and numbers such as production or us-east-1. Register anything else with add_secret().
• Well-known key formats: OpenAI, Anthropic, AWS, GitHub, Google, Slack, JWT and Bearer tokens, and credential query parameters such as ?key=, plus every parameter the OpenTelemetry semantic conventions say to redact (X-Amz-Signature, X-Amz-Credential, X-Amz-Security-Token, X-Goog-Signature, AWSAccessKeyId, Signature, sig).

It covers span names, attributes, exception events, status messages, link attributes, resource attributes, log bodies and log attributes. The environment is read once per export batch, so a secret set while the app runs is caught from the next batch on.

A filter can't recognize a secret that matches none of these rules, such as an unusual password a user types. That's why the two go together: the filter is the fix, and pytest --traceleak is the proof.

Any language: traceleakcheck run

The checker speaks OTLP, the format every OpenTelemetry SDK exports, so it can check an app in any language from the outside. No pytest is needed. It's tested with Node and Python apps, and any SDK that reads the standard OTEL_EXPORTER_OTLP_* settings should work. For exporters that use gRPC, which is common in Go and .NET, install traceleakcheck[grpc].

traceleakcheck run --secret OPENAI_API_KEY -- node app.js

It runs these steps:

1. Starts the receiver.
2. Sets the app's OPENAI_API_KEY to a fresh fake key.
3. Points the app's OpenTelemetry SDK at the receiver with the standard OTEL_* settings.
4. Runs the app until it exits.
5. Prints the same plain-English report. The exit code is 1 on a leak and 2 if it couldn't check, so CI fails.

Servers. Pass a command that exercises the app. {OPENAI_API_KEY} in it is replaced with the fake key, and that works in every shell:

traceleakcheck run --ready-url http://localhost:3000/health \
  --drive "curl -s localhost:3000/chat?q={OPENAI_API_KEY}" -- node server.js

Option	What it does
--secret ENV[=KIND]	Env var to fill with a fake secret: openai (default), anthropic, azure, aws, github, jwt or password. Repeatable.
--drive CMD, --ready-url URL	Treat the app as a server: wait until it answers, run CMD, then stop it, including child processes.
--expect traces,logs	Signals that must arrive, or the result is "couldn't check" (default: traces).
--html, --report, -v	HTML report, JSON report, exact field paths.

The control is weaker than in pytest. traceleakcheck can't send a marker from inside another process. It checks that telemetry arrived on every expected signal and prints how many spans, logs and metrics it scanned. An app that exits without flushing loses its last batch, so make sure it shuts the SDK down (Node: await sdk.shutdown()). See examples/node_agent for a leaking Node app, and the Action's command: input for CI.

The fix for any language is an OpenTelemetry Collector between your app and your backend. examples/collector/redact.yaml removes well-known key formats and credential query parameters from span names, attributes, error events, status messages, log bodies and attributes, metric attributes, and resource attributes.

To prove it covers your app, run the app through the Collector and the Collector into traceleakcheck. Start the Collector (otelcol-contrib) with the receiver's port in TLC_PORT, then:

TLC_PORT=4400 otelcol-contrib --config examples/collector/redact.yaml &
traceleakcheck run --port 4400 --app-endpoint http://127.0.0.1:4318 -- node app.js

On Windows, examples/collector/prove.ps1 does both steps against the Node demo's --proof mode. That mode leaks the key through every channel above, and also sends a password with no known shape in a ?token= URL through each of them, so only the query-parameter rule can catch it. Without the Collector it leaks in 15 places; with it, the result is ✅ clean. CI runs this proof on every push with otelcol-contrib 0.164.0.

Tested on real code

We pointed TraceLeakCheck at the OpenTelemetry Demo's AI shop assistant (LangGraph + Traceloop), unmodified, talking to a fake LLM:

Scenario	Without the filter	With scrub()
A shopper pastes a key into the chat	🚨 leaked in 9 places	✅ clean
The service's own API key	✅ clean	✅ clean
The LLM gateway rejects the key and echoes it	🚨 leaked in 15 places	✅ clean

One secret becomes many copies, because every layer of an agent (chat step, agent step, workflow) records its own copy. The run also surfaced a real integration bug in TraceLeakCheck: the Traceloop SDK ignores OTEL_EXPORTER_OTLP_*. That's now fixed.

Survey: do GenAI instrumentations leak secrets?

Read the survey as a page, with the full results matrix.

survey/FINDINGS.md runs the OpenAI SDK under OTel openai-v2, OpenLLMetry, OpenInference, Logfire and httpx instrumentation, across 8 leak scenarios:

• No content opt-out we tested stops a secret in an error message. That includes TRACELOOP_TRACE_CONTENT=false, OPENINFERENCE_HIDE_* and NO_CONTENT.
• ?key= query auth is exported verbatim in http.url.
• Logfire's default scrubbing skips LLM messages and exception messages, even when they contain "API key".
• OTel openai-v2's documented legacy CAPTURE_MESSAGE_CONTENT=true falls back to NO_CONTENT on util-genai 1.x, with only a log warning.
• Streamed replies are reassembled, so a secret split across chunks still leaks. But OTel openai-v2 never ends the span of an abandoned stream, so its contents are never exported. traceleakcheck now reports that as incomplete instead of a false clean.

Limits

• Any language via traceleakcheck run, but with a weaker control than the pytest plugin, and no detection of spans left open. scrub() is Python-only; other languages use the Collector config.
• Sampling is forced on with OTEL_TRACES_SAMPLER=always_on. An app that passes an explicit sampler in code can still drop the span that leaks.
• Telemetry that isn't OTLP, such as metrics read by a Prometheus reader, can't be checked. Skip it with --traceleak-ignore=metrics.

Compared to Weir

Weir uses the same canary idea, and it covers more ground: it runs agent flows against a fake model and checks console output and network requests as well as telemetry. TraceLeakCheck focuses on the OTLP export boundary, needs no fake model, and ships a fix (scrub()) alongside the check.

Weir's README says it "finds whole values only, not hashed, split or truncated ones". TraceLeakCheck catches truncated secrets. It also has tests for base64 at any byte offset, secrets used as attribute keys, protobuf fields the decoder drops, and mixed-case hex. As of 2026-11-01, Weir's README doesn't say either way whether it handles these.

Library API

from traceleakcheck import OTLPReceiver, make_canary, scan

with OTLPReceiver() as rx:
    canary = make_canary("openai")
    # point your OTLP exporter at rx.endpoint, run your app with canary.value as the API key...
    findings = scan(rx.snapshot(), [canary])

Develop

uv sync
uv run pytest