pgfuzz secrets is the secret-scanner sibling of pgfuzz pii. Same statistical question, different corpus: “of 100 plausible-looking AWS access keys embedded in code-snippet payloads, how many does my live Secret Scanner guardrail catch?”

Quick start

pgfuzz secrets \
   --target https://gateway.your-domain.com/api/$PG_UUID/v1/chat/completions \
   --token pg_live_… \
   --count 200

The Secret Scanner guardrail must be enabled and in block mode.

Flags

Flag	Default	Notes
--target	(required)	Full chat-completions URL of the gateway
--token	(required)	Bearer token (pg_live_…) — must have the chat scope
--model	openai:gpt-4o-mini	Model identifier the gateway should route to
--count	100	Total synthetic payloads to fire, distributed across kinds
--kinds	aws,github,slack,openai,anthropic,stripe,jwt	Secret kinds to generate
--concurrency	16	Parallel HTTP workers
--timeout	30	Per-request timeout (seconds)
--report	tty	Output format: tty (table) or json

Kinds

Each generator emits a syntactically-valid but locally-generated string — matches the published prefix / format, never deployed anywhere real. PromptGate’s scanner matches on shape, so synth strings trip it identically to real credentials.

Kind	Prefix	Length / shape
aws	AKIA	+ 16 chars [A-Z2-7]
github	ghp_	+ 36 chars alphanumeric
slack	xoxb-	+ 11 digits + - + 11 digits + - + 24 alphanumeric
openai	sk-	+ 48 alphanumeric
anthropic	sk-ant-api03-	+ 95 chars (alphanumeric + _-)
stripe	sk_live_	+ 24 alphanumeric
jwt	(no prefix)	3 base64url-ish segments separated by .

Carrier templates

The raw secret is wrapped in plausible code-snippet payloads — the way an LLM agent typically leaks one:

Here's the snippet I'm working with — please refactor it:
```const key = "AKIAIOSFODNN7EXAMPLE";```

Why is my script failing? I set the env var like this:
export AWS_KEY="AKIAIOSFODNN7EXAMPLE"

The README says to put this in config.json:
{ "token": "AKIAIOSFODNN7EXAMPLE" }

Each template variant is chosen at random per sample.

Verdict

Same semantics as pgfuzz pii:

Outcome	Condition
caught	HTTP 422 + body mentions a secret kind (AWS, GitHub, Slack, OpenAI, …) or secret / blocked / redact
missed	HTTP 2xx
error	Anything else — excluded from recall

Output

TTY

pgfuzz — recall report
  target:   https://gateway.your-domain.com/api/.../v1/chat/completions
  total:    210
  duration: 2.9s
  p50/p95:  58ms / 142ms

KIND       VARIANT  TOTAL  CAUGHT  MISSED  ERRORS  RECALL
----       -------  -----  ------  ------  ------  ------
anthropic  inline   30     30      0       0       100.0%
aws        inline   30     30      0       0       100.0%
github     inline   30     30      0       0       100.0%
jwt        inline   30     27      3       0       90.0%
openai     inline   30     30      0       0       100.0%
slack      inline   30     30      0       0       100.0%
stripe     inline   30     30      0       0       100.0%
----       -------  -----  ------  ------  ------  ------
ALL        (all)    210    207     3       0       98.6%

JSON

Same payload shape as pgfuzz pii — full Summary with run, cells[], overall.

Examples

Verify all kinds before a release

pgfuzz secrets --target $URL --token $TOKEN --count 350
# expect recall ≥ 95% per kind

Catch a regression after a scanner-pattern change

pgfuzz secrets --target $URL --token $TOKEN --count 1000 \
   --kinds aws,github,openai --report json \
   | jq '.cells[] | select(.recall < 0.99)'

Per-kind deep dive

pgfuzz secrets --target $URL --token $TOKEN --count 200 \
   --kinds jwt --concurrency 8

JWTs are the most format-flexible kind (any base64-ish 3-segment value matches). Use this to tune your false-positive vs recall tradeoff.

---

Next: Roadmap — Phase 2 subcommands.

---

© Akyros Labs LLC. All rights reserved.