Skip to content

Reference

Detector reference

Every core Vigil detector is deterministic: the same input always produces the same verdict. An optional AI-assisted injection check can be turned on as an extra layer; it only runs when the patterns did not already flag the prompt. Here is what each one looks for — and what it does not catch.

Prompt injection signals

Each matching signal adds weight; weights combine into a 0–100 risk score compared against your rule's threshold.

Signal IDNameLooks for
override_instructionsInstruction overrideAsks the model to ignore, forget or override previous or system instructions.
system_prompt_extractionSystem prompt extractionAsks the model to reveal or repeat its system prompt or hidden instructions.
safety_bypassSafety bypassAsks to disable or circumvent safety rules, filters or guardrails.
jailbreak_personaJailbreak personaKnown jailbreak personas such as DAN or "developer mode".
no_restrictionsRestriction removal"Pretend you have no rules" style framing.
role_markersChat role marker injectionFake chat-template markers such as [INST] or system: lines.
data_exfiltrationData exfiltrationInstructions to send data to an external URL, including markdown image beacons.
new_instructionsInjected instruction blockText that announces "new instructions:" or a replacement task.
encoded_payloadEncoded payloadRequests to decode long base64 or rot13 blobs.
persona_switchPersona switch"You are now…" role reassignment.

Personal data (PII)

Card numbers are checked with the Luhn algorithm and IBANs with their checksum to reduce false positives.

  • email — Email address
  • phone — Phone number
  • ssn — US Social Security number
  • credit_card — Payment card number
  • iban — IBAN
  • ip_address — IP address

Secrets and credentials

  • AWS access key
  • OpenAI API key
  • GitHub token
  • Slack token
  • Stripe secret key
  • Google API key
  • Vigil ingest key
  • Private key block
  • JSON Web Token
  • Credential assignment (password=, api_key=…)

Tool allowlists

Tool calls are checked by name against the agent's allowlist before they run. When argument inspection is on, string values anywhere in the arguments are also scanned for destructive shell commands, remote download-and-execute, destructive or unbounded SQL, SQL injection, path traversal and sensitive files, cloud metadata endpoints and prompt injection. You can also pin an MCP tool's description fingerprint; if an agent later reports a different description for that tool (a sign of tool poisoning), the guardrail triggers. Argument checks are pattern-based and do not understand every tool's semantics — keep allowlists tight.

Known limits

Pattern-based detection misses novel phrasings, other languages and heavily obfuscated attacks. That is why we pair detection with hard tool boundaries and fail-closed defaults — read our manifesto.