Orca AI Incident Archive
An open, sourced database that separates AI agent incidents with real victims from research demos, one Markdown record at a time.
- Source repo
- Continuum-AI-Corp/Orca-AI-Incident-Archive
- Stars
- ★ 329
- Last updated
- 9d ago
- License
- NOASSERTION
- Primary language
- HTML
- FA score
- 60/100 · Some gaps
At a glance
- Works with
- Universal · cross-platform
- You'll need
- Typical use
- A security researcher needs clickable primary evidence for the claim that an agent actually caused damage, and filters real_harm: true across the 153 qualifying records.
- Main limitation
- Records are written in English; titles and summaries exist in seven languages, but complete translated text only exists under incidents/i18n/zh/.
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-12+28.3 ★/day over 7 days
- Source review
- 60/100 · Some gaps
What does this agent do, and when should you use it?
Orca AI Incident Archive is an open dataset of real-world AI agent security events spanning 2025-01 through 2026-11-08, with 417 records plus one precursor traceable to 2024-12-01. Each record is a Markdown file with a YAML header, an attack-chain diagram and at least one clickable primary source, and each is forced to answer three fields first: real_harm (is there a confirmed victim), ai_involvement (did a primary source confirm AI was involved), and kind (incident, vulnerability disclosure, research demo, threat report or policy move). The repository ships month-by-month directories under incidents/, seven attack-surface topics, regional slices, a SCHEMA.md field spec and a taxonomy/, plus dist/ with JSON, CSV, statistics and every source URL. A single index.html works offline in seven languages. Data is released under CC BY 4.0 with attribution.
The repository organizes four bodies of content: per-month Markdown records under incidents/ (YAML header, attack-chain diagram, primary source links), attack-surface indices under topics/, regional slices under regions/, and full Chinese text of every record under incidents/i18n/zh/. Each record cites external primary sources to back its judgments, and dist/ exports JSON, CSV, aggregate statistics and a complete list of source URLs for downstream analysis. index.html is a single-file front end that reads this data for offline browsing, with titles and summaries in English, Chinese, Japanese, Korean, German, French and Spanish. docs/data-quality.md logs every change made across four verification rounds, including correcting PaperCut's "domain admin in six hours" to seven minutes and downgrading Step Finance to grade D because the primary reporting never mentions AI. Records carry ids such as orca:2026-07-09-openai-agents-breach-huggingface for single-record citation.
- A security researcher needs clickable primary evidence for the claim that an agent actually caused damage, and filters real_harm: true across the 153 qualifying records.
- An AI governance or compliance team builds an internal risk register from the 49 critical entries, using the severity criteria in taxonomy/severity.md as a baseline.
- A threat-intelligence analyst consumes dist/ JSON and CSV to correlate supply-chain events such as Shai-Hulud, TanStack and CHAINDROP into a single timeline.
- An incident responder browses topics/ for prompt-injection, MCP and credential-theft cases to compare against their own agent architecture's exposure.
- A journalist or technical writer cites a single record by its orca: id while preserving any disputed or unverified label attached to it.
- A data engineer opens index.html offline to walk reviewers through monthly and regional charts without network access.
How do you install or deploy this agent?
This repository is a dataset plus a static site; there is no server or SDK to build. Cloning is the whole installation:
How do you use this agent?
Open the repository root's index.html in a browser for offline, multilingual browsing. For programmatic use, consume the files under dist/ directly, for example filtering on real_harm with jq:
What are this agent's strengths and limitations?
- The mandatory three fields (real_harm, ai_involvement, kind) keep actual damage distinct from proof-of-concept research instead of inflating a headline count.
- No source, no entry: 417 records carry 868 source links across 795 unique URLs, with zero unsourced records.
- Uncertain claims are labelled rather than deleted, so ai_involvement: unverified and disputed states stay searchable alongside the rebuttal.
- Ships both structured exports in dist/ (JSON, CSV, stats) and a single offline index.html, so it works as data and as a browsing tool.
- CC BY 4.0 with a ready-made BibTeX entry makes citation and redistribution straightforward.
- Records are written in English; titles and summaries exist in seven languages, but complete translated text only exists under incidents/i18n/zh/.
- Severity and classification are editorial judgments, not official vendor or regulator findings, so they need independent verification before audit use.
- Scope is limited to publicly disclosed events and excludes undisclosed vulnerability detail and exploit code, so it cannot be used for reproduction or active testing.
- There is no documented automated sync or subscription feed; citations must account for the 2026-11-08 build date.
- The fourth verification round still found roughly 11% missing, so the dataset is not an exhaustive statistical census.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Orca AI Incident Archive This agent | 60 · Some gaps | — | ★ 329 | 9d ago | HTML | — |
| Prismor | 61 · Some gaps | CLIFree + model costs | ★ 583 | today | Python | Codex · Claude Code · OpenAI API · Claude API |
| HOL Guard | 80 · Good | CLIFreemium | ★ 1.1k | today | Rust | Codex · Claude Code |
| SkillSpector Security Scanner | 75 · Good | CLIFree + model costs | ★ 27k | today | Python | Codex · Claude Code · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
This is a data archive rather than an executing agent, so the privilege surface is inherently small: the CI workflow declares permissions: contents: read, matching least privilege (2). There is no user-confirmation mechanism; contribution/merge relies on prose rules only (1). Data-flow transparency is fair: README explains Markdown+YAML records, dist/ JSON/CSV, and an offline index.html (2). Sensitive data: it states only publicly disclosed events are recorded, with no undisclosed vulnerability detail or exploit code, and linked sources remain their owners' copyright (2). External effects: static content with no destructive defaults (2). Rollback: only 'corrections go into the record, never silently over it'; no versioned rollback or revert mechanism (1). Dependency security: CI uses unpinned third-party actions (checkout@v4, setup-python@v5) with no dependency lock or vulnerability scanning visible (1). Source attribution: every record needs a clickable primary source, 0 records without source, CC BY 4.0 attribution required — thoroughly handled (3).
Self-consistency: badge and body numbers largely agree (417 records, 153 confirmed harm, 795 unique URLs), but the badge says primary_sources 795 URL while the quality table says 868 links across 795 unique URLs — minor wording mismatch (2). Dependency availability: only Python 3.12 and scripts are referenced; no requirements or lock file, and link checking is advisory and samples only 60 links (1). Failure messages: CI emits a clear error and diff when derived files are out of sync, but validate.py/build.py are not shown, so their runtime error messaging cannot be assessed (1).
Audience and scenarios: the 'I want to…' navigation table covers chronological, by-surface, by-region, data-analysis and offline browsing — clear audience targeting (3). Capability boundaries: explicit inclusion criteria and out-of-scope list (pure LLM content safety, unrelated vulns, claims without primary source) (2). Trigger precision: as a data archive there is no trigger, but the four 'what counts as a record' rules are qualitative and edge cases depend on editorial judgement (1). Environment fit: offline single-file index.html, seven languages, multi-format dist/ — good fit; but no runtime/dependency instructions (2).
Information architecture: README, SCHEMA, taxonomy, topics, regions, docs, dist are cleanly layered with good navigation (3). Install notes: no install/usage steps beyond an implied 'python scripts/build.py'; prerequisites and dependencies undocumented (1). Naming stability: record ids use orca:date-slug and are used for citation, fairly stable, but no formal naming spec (2). Examples and FAQ: citation example, contribution rules and issue templates exist, but no dedicated FAQ (2). Known limitations: README admits ~11% coverage gap, editorial-judgement classification, and non-official status (2). License: LICENSE is full CC BY 4.0, but repository metadata says NOASSERTION — metadata/file mismatch (2). Versioning/changelog: only 'Built 2026-11-08' and four verification rounds; no version number or CHANGELOG (1). Maintenance responsibility: publisher is unverified by FollowAgents and identity unknown; no maintainer list or governance (1).
Output usability: structured Markdown+YAML records, dist/ JSON/CSV, offline index.html — directly usable for analysis (2). Marginal value: the real_harm/ai_involvement/kind triad distinguishes actual harm from research demos, a clear increment over generic incident lists (2). Cost-benefit: static data, no execution needed, low acquisition cost; but accuracy depends on editorial judgement and coverage is self-admittedly incomplete (2).
Claim traceability: every record requires at least one clickable primary source, 0 without source, 359 grade-A primary sources, citations carry ids (3). Cross-source corroboration: 16 records are flagged disputed with both accounts preserved, but most records appear single-source with no systematic multi-source cross-check (2). Fact/inference separation: kind/ai_involvement labels and a 'classification is editorial judgement' disclaimer exist, but severity and classification remain mixed into records, so separation is incomplete (2).
- Repository license metadata is NOASSERTION while LICENSE states CC BY 4.0 — confirm before reuse.
- Publisher identity is unverified by FollowAgents; maintenance responsibility and update path are unclear.
- CI uses third-party actions not pinned to SHAs, with no dependency lock or vulnerability scanning visible.
- Data accuracy depends on editorial judgement; README admits ~11% coverage gap and some records are single-source.
- No install/run instructions or version/changelog; reproducing the build requires inference.
FAQ
Is Orca AI Incident Archive open source?
Is Orca AI Incident Archive safe to use?
Is Orca AI Incident Archive still maintained?
Who makes Orca AI Incident Archive?
Thanks — we use this to improve the page.