AB sentinel-custom-provider-check
Verifies that Sentinel's PII/secret redaction guardrail survives routing through CUSTOM_PROVIDERS to a self-hosted OpenAI-compatible backend (vLLM, TGI, Ollama, or anything speaking /v1/chat/completions) — not just Sentinel's first-party Anthropic/OpenAI/Google routes. Starts the local Sentinel gateway container with CUSTOM_PROVIDERS pointed at the backend, sends a synthetic-PII probe straight through it, and reads the backend's own raw response to confirm the guardrail actually fired on that hop. Use when: user says "check if Sentinel redacts through my custom model", "verify Sentinel guardrails on my self-hosted backend", "test CUSTOM_PROVIDERS redaction", or "does my vLLM/Ollama server see PII".
Verifies that Sentinel's PII/secret redaction guardrail survives routing through CUSTOMPROVIDERS to a self-hosted OpenAI-compatible backend (vLLM, TGI…
As a process B 69/100 · Nearly there — weak spots: result and completion, inputs and preconditions, progress reporting
What is at stake
Medium-severity findings: the skill is probably honest, but read what alarmed the scanner.
Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.
The instructions or scripts send data out: environment variables, keys, file contents, chat history. You may never notice, because the agent performs the upload as "part of the task".
If the upload is not needed for the task, remove it: catalogs flag such skills and delist them. If it is needed, name the destination explicitly, say what leaves the machine, and give the user a switch.
How to improve
- Your own cases (evals/evals.json, 4–6 real requests with expected answers): the full check would then run those instead of a model-drafted suite.
- A spec.yaml with trigger phrases and assertions — a behaviour contract for CI; `skilltest init` writes a template.
Guard findings · 2
✓ No critical or high findings
Medium and low: 2
-
medium Exfiltration
net-redirectable-api-keyskill.py:89Helper sends the API key to a host configured by an environment variable — the key can be redirected to another serverAPI key + configurable base URL from environment
-
medium Exfiltration
exfil-webhook-urlskill.py:194Webhook / callback URL commonly used for exfiltration (verify the destination) (quoted — discussed, not commanded)f"https://api.telegram.org/bot{token}/sendMessage",quoted
Files scanned: 4. Evidence is masked. Grey chips explain why severity was lowered.
Against the Agent Skills spec
✓ No remarks against the Agent Skills spec
Process rating: all ten parameters 69/100
- 0Result and completion. Does not say what the result is
- 0Inputs and preconditions. Does not say what the process needs to start
- 0Progress reporting. Says nothing while it works
- 70When it triggers. States when to use, but not when not to
- 100Tools and files. No external tools needed
- 100Steps. 17 steps
- 100Failures and branches. 1 branches, has a failure section
- 100Consistency. Name and required fields are in place
- 100Execution cost. Instruction body is 1357 tokens
- 100Running it twice. Mutating operations check current state
Everything here is measured from the skill text rather than judged by a model, so the numbers are checkable. A parameter weighs more when it is a more common reason for the process to stall.
Quality signals
- +4Description does not say when NOT to use the skill (false activations)
- +3Output format is not stated: the model decides each time
- +1No license
- +2Single-language instructions
- +5Description quotes 4 example trigger phrases
- +3Description length 707: enough signal without eating the budget
- +4Structure: 9 headings
- +3Step-by-step instructions: 17 items
- +4Has examples (3 code blocks)
Quality base 70; lint remarks subtract, signals add up to 100. Result: 89.