LLM Security Red-Team Evidence Packet: What to Capture Before a Decision

The direct answer

Assemble reviewable evidence for a scoped adversarial campaign against an LLM application without turning assumptions or producer claims into proof. The working output is A claim-to-evidence packet with provenance, freshness, contradiction, and NOT_TESTED fields.

For LLM Security Red-Team, the bounded capability is a scoped adversarial campaign against an LLM application. Begin only when the team can supply authorized application access, scope documentation, prohibited actions, test data, and incident contacts. The documented delivery target is a threat model, per-attack evidence record, and prioritized control list; anything broader requires a new scope and a new authority decision.

The copyable evidence packet

This evidence packet is for teams that need attack evidence, not only a design checklist. It begins with authorized application access, scope documentation, prohibited actions, test data, and incident contacts and stays inside the documented workflow: authorization, threat modeling, attack-case selection, safe execution, per-attack evidence, control mapping, prioritization, and retest planning. For LLM Security Red-Team, the evidence packet remains reviewable because its decisions have named owners, evidence fields, and stop conditions.

The LLM Security Red-Team evidence packet keeps each claim separate from its source and from the reviewer decision that accepts or rejects it. For this evidence-packet task, a URL or file name supplies provenance but not automatic proof; an executor saying “done” remains an assertion until a separate observation supports the exact criterion.

Machine-readable evidence row

{
  "claim_id": "CLM-01",
  "claim_type": "OBSERVED",
  "content": "Scope and prohibited actions are signed off.",
  "evidence": "attach a direct readback or test receipt",
  "provenance": {
    "source": "named path or system",
    "observed_at": "ISO-8601",
    "method": "inspection or test"
  },
  "status": "NOT_TESTED",
  "contradictions": [],
  "reopen_if": "testing without written authorization"
}

Contradiction rule

When two admissible records disagree, retain both and mark the claim REVIEW. Do not average incompatible observations or choose the convenient one. The security observer records what changed, which evidence applies to the current boundary, and what must be rerun. When required evidence is unavailable, the status stays NOT_TESTED.

Run the workflow as a sequence of decisions

The LLM Security Red-Team evidence packet follows this working sequence: authorization, threat modeling, attack-case selection, safe execution, per-attack evidence, control mapping, prioritization, and retest planning. Within this artifact, each phrase marks a state boundary for a scoped adversarial campaign against an LLM application. A stage output becomes the next named input, while a failed, missing, or unavailable check keeps the dependent evidence packet decision closed.

Separate failure register

• FAIL-01: Testing without written authorization.
• FAIL-02: Attack cases copied from a checklist without system context.
• FAIL-03: Successful prompts recorded without downstream impact evidence.
• FAIL-04: Unsafe data or tools left in scope.
• FAIL-05: Controls recommended without a retest condition.

The register supplies negative cases for the complete acceptance set. A reviewer determines affected checks from observed evidence; array position never asserts that one failure proves or disproves one criterion.

The producer can explain what it attempted, but the security observer evaluates the evidence. If the artifact changes, its prior verdict expires. This is especially important for a scoped adversarial campaign against an LLM application, where a plausible narrative can hide a stale configuration, an untested negative case, or an authority mismatch.

Failure and recovery drills

A useful LLM Security Red-Team evidence packet explains what happens when its happy path breaks. These drills come from the accepted product truth record rather than a claim that every buyer has each failure. Use safe synthetic or authorized observations for a scoped adversarial campaign against an LLM application, and keep private credentials out of every fixture.

1. Testing without written authorization.

Detect for LLM Security Red-Team: system owner captures a direct readback or safe fixture that makes this evidence packet condition observable. Its record binds source, time, method, and the current ART-10-02 fingerprint.

Contain the evidence packet: stop only the affected LLM Security Red-Team path after observing “testing without written authorization”. Preserve its failed material and last verified state instead of erasing evidence or blindly repeating an external effect.

Recover and prove: apply the smallest authorized LLM Security Red-Team correction, then have a distinct reviewer re-evaluate the complete accepted check set. Do not select one check merely because it shares this failure's list position. If any affected evidence packet check cannot run, its result remains NOT_TESTED.

2. Attack cases copied from a checklist without system context.

Detect for LLM Security Red-Team: red-team lead captures a direct readback or safe fixture that makes this evidence packet condition observable. Its record binds source, time, method, and the current ART-10-02 fingerprint.

Contain the evidence packet: stop only the affected LLM Security Red-Team path after observing “attack cases copied from a checklist without system context”. Preserve its failed material and last verified state instead of erasing evidence or blindly repeating an external effect.

Recover and prove: apply the smallest authorized LLM Security Red-Team correction, then have a distinct reviewer re-evaluate the complete accepted check set. Do not select one check merely because it shares this failure's list position. If any affected evidence packet check cannot run, its result remains NOT_TESTED.

3. Successful prompts recorded without downstream impact evidence.

Detect for LLM Security Red-Team: security observer captures a direct readback or safe fixture that makes this evidence packet condition observable. Its record binds source, time, method, and the current ART-10-02 fingerprint.

Contain the evidence packet: stop only the affected LLM Security Red-Team path after observing “successful prompts recorded without downstream impact evidence”. Preserve its failed material and last verified state instead of erasing evidence or blindly repeating an external effect.

Recover and prove: apply the smallest authorized LLM Security Red-Team correction, then have a distinct reviewer re-evaluate the complete accepted check set. Do not select one check merely because it shares this failure's list position. If any affected evidence packet check cannot run, its result remains NOT_TESTED.

4. Unsafe data or tools left in scope.

Detect for LLM Security Red-Team: data owner captures a direct readback or safe fixture that makes this evidence packet condition observable. Its record binds source, time, method, and the current ART-10-02 fingerprint.

Contain the evidence packet: stop only the affected LLM Security Red-Team path after observing “unsafe data or tools left in scope”. Preserve its failed material and last verified state instead of erasing evidence or blindly repeating an external effect.

Recover and prove: apply the smallest authorized LLM Security Red-Team correction, then have a distinct reviewer re-evaluate the complete accepted check set. Do not select one check merely because it shares this failure's list position. If any affected evidence packet check cannot run, its result remains NOT_TESTED.

5. Controls recommended without a retest condition.

Detect for LLM Security Red-Team: remediation owner captures a direct readback or safe fixture that makes this evidence packet condition observable. Its record binds source, time, method, and the current ART-10-02 fingerprint.

Contain the evidence packet: stop only the affected LLM Security Red-Team path after observing “controls recommended without a retest condition”. Preserve its failed material and last verified state instead of erasing evidence or blindly repeating an external effect.

Recover and prove: apply the smallest authorized LLM Security Red-Team correction, then have a distinct reviewer re-evaluate the complete accepted check set. Do not select one check merely because it shares this failure's list position. If any affected evidence packet check cannot run, its result remains NOT_TESTED.

Ownership and handoff

For this LLM Security Red-Team evidence packet, the adjudication role is security observer. That role judges frozen acceptance evidence for a scoped adversarial campaign against an LLM application without becoming the product owner, legal adviser, security authority, or buyer. Its handoff retains open gaps, failed evidence, changed hashes, and the next action permitted for ART-10-02.

Evidence and acceptance

Use these product-specific statements as candidate acceptance checks:

• Scope and prohibited actions are signed off.
• Attacks map to named threat hypotheses.
• Each result has reproducible evidence.
• Findings separate exploitability from impact.
• Control fixes have retest cases.

For every LLM Security Red-Team evidence packet check, retain the tested object, environment or source, observation time, method, expected result, actual result, verifier identity, and artifact hash. In this ART-10-02 record, label a direct readback OBSERVED, a reproducible transformation COMPUTED, and an interpretation JUDGMENT; never merge those states into one confident claim.

The admitted Search Console packet contained no article-specific demand observation for this exact topic. The page is therefore justified by its distinct operator job and product truth, not by an invented volume estimate. Performance remains unknown until measured after an authorized release.

The product boundary remains controlling: A scoped campaign cannot certify the system, prove the absence of unknown vulnerabilities, or replace broader application and infrastructure security testing.

Implementation checklist

• The evidence packet names the distinct reader job: Assemble reviewable evidence for a scoped adversarial campaign against an LLM application without turning assumptions or producer claims into proof.
• The input boundary is explicit: authorized application access, scope documentation, prohibited actions, test data, and incident contacts.
• The intended deliverable is explicit: a threat model, per-attack evidence record, and prioritized control list.
• Every required acceptance check has current evidence or an honest NOT_TESTED status.
• At least one negative fixture covers testing without written authorization.
• The security observer is distinct from the artifact producer.
• Rollback or reopen conditions are written before consequential action.
• No ranking, traffic, conversion, compliance, certification, or buyer-outcome guarantee was added.

When this LLM Security Red-Team evidence packet has a failed item, repair that named item and rerun its dependent checks. Keep the frozen threshold intact; the remaining checks cannot establish that the failed ART-10-02 condition probably holds.

Related campaign guides

These four crawlable links cover the other LLM Security Red-Team operator artifacts. Each leads from the current evidence packet to a different reader task and promised artifact, rather than to a keyword variation of this page.

• LLM Security Red-Team Input Contract: Required Fields, Rejection Rules, and Handoff
• LLM Security Red-Team Operator Runbook: Daily Checks, Escalations, and Handoff
• LLM Security Red-Team Control Dashboard: Signals, Alerts, and Review Cadence
• LLM Security Red-Team Change Protocol: Versioning, Canary Tests, and Rollback