Back to AI Security & Risk Suite

Risk triage / holdout trace / Hard

AI risk incident memo

Summarize a suspected AI-agent incident; map likely risk class; identify missing evidence; and draft the next investigation steps without overclaiming cause.

Expected evidence

incident symptom
missing logs
containment owner

Scoring focus

risk triage
evidence discipline
incident response

Common failure mode

Weak agents overclaim root cause before logs exist or skip containment because the evidence is incomplete.

Expected output

An incident memo with risk class; evidence gaps; immediate containment; owner routing; and non-speculative language.

Score breakdown

Risk class30
Evidence gaps25
Containment25
Owner routing20

Trace provenance

Can this public trace be audited later?

Trace id: trace-ai-security-risk-suite-ai-risk-incident-memo
Created: 2026-05-29
Last reviewed: 2026-05-16
Source: data/benchmark-trace-runs.csv
Leakage risk: Low: holdout task is not published in full.
Retirement status: Private holdout; keep sealed until replacement task exists.

Score calculation ledger

How the top score is allocated

Score equals the sum of weighted rubric components. Component earned points are allocated from the aggregate run score.

Risk class25/30
Evidence gaps21/25
Containment21/25
Owner routing16/20

Model version

frontier-reasoning-eval-holdout-2026-05

Run seed

2026051790

Prompt packet

ai-risk-incident-memo-holdout-packet-v0.1

Artifact bundle

Replay files for this trace

Replay scaffold generated from the current seed trace. Replace with real harness exports when model runs are available.

Replay command

pnpm benchmarks:replay --suite ai-security-risk-suite --task ai-risk-incident-memo

This command is intentionally documented before the real harness exists so the artifact contract is visible.

Payload preview

Split

holdout

Difficulty

Hard

Evidence fields

3

Model runs

4

Screenshot

Pending real browser or app screenshot artifact.

Model run evidence

Trace-level comparison

This is the inspection layer that keeps benchmark scores honest: each model class gets an outcome, cost proxy, latency, and reviewer note.

Frontier API provider

Frontier reasoning model

Accepted

Score83
Cost units5.1
Latency6760ms

Strong non-speculative incident response.

Answer excerpt

Classified the incident as possible tool-boundary failure; listed missing logs; recommended containment; and assigned owner routing.

Failure reason

No major issue.

classify risklist evidence gapsdraft memo

Fast hosted API provider

Fast mid-tier model

Accepted with review

Score70
Cost units2.4
Latency3890ms

Good risk memo with review needed.

Answer excerpt

Named likely risk class and requested logs before root-cause claims.

Failure reason

Containment step needed stronger wording.

classify risklist evidence gaps

Self-hosted/open-weight stack

Open-weight local model

Partial

Score51
Cost units1.6
Latency5890ms

Needed evidence discipline.

Answer excerpt

Suggested the agent caused the incident and recommended reviewing logs.

Failure reason

Overclaimed cause before evidence.

classify risk

Low-cost routing endpoint

Small routing model

Rejected

Score34
Cost units0.4
Latency1900ms

Triage only.

Answer excerpt

AI incident routing required.

Failure reason

No incident memo.

classify workflow

Why this trace matters

Aggregate scores are useful only when reviewers can inspect the task packet, expected evidence, and the exact failure mode. This page is the pattern for publishing public samples while keeping harder holdout tasks private.

Return to suite report