An open, structured dataset (an incident database) of publicly documented security incidents, vulnerability disclosures and threat reports involving LLM applications and AI agents, for security engineers, researchers and policy teams who need evidence about prompt injection, data exfiltration, supply-chain compromise and agent misuse rather than headlines. Every event is one schema-validated JSON file, coded under a written codebook by the role AI plays in it (weapon, target or surface), how the attack got in, what authority the AI held, how the effect got out, and what harm resulted, and cross-referenced to the OWASP Top 10 for LLM Applications, the OWASP Top 10 for Agentic Applications and MITRE ATLAS.
Browse it at https://basitalisandhu.github.io/ai-agent-incidents/ (searchable table, one page per event, RSS feed, JSON, JSONL, CSV, stats).
Security discussion treats "AI" as one problem. The public record says it is at least three: AI as a weapon that lowers the cost of established attacks, AI as a target whose data, prompts, credentials and serving stack can be extracted or disrupted, and AI as an attack surface, because organisations now wire instruction-following models to privileged tools, private data and untrusted content at once. Reasoning about which controls matter, and arguing with vendors and auditors about them, needs evidence that is structured, sourced and reproducible, not a list of headlines. Until now no such dataset existed in a form that can be queried, diffed, validated in CI and cited.
The seed data and the coding scheme come from the paper AI as Weapon, Target, and Surface: A Threat Taxonomy and a Deterministic Control Plane for Securing LLM Agents (Muhammad Basit Ali, 2026); its codebook is reproduced in docs/codebook.md. This repository turns that dataset into a maintained, contributor-friendly record.
- Security engineers and architects deciding which controls an agent deployment
needs: filter by
authorityandchannel_outto see what actually caused harm. - Researchers who need a coded, citable sample of real events, with a codebook and a reproducible build.
- Vendors and policy teams mapping incidents to OWASP and ATLAS for risk registers, threat models and compliance narratives.
- Journalists and educators who want the facts of each event with its primary source, in one place.
- Which AI agent security incidents involved prompt injection, and what did the attacker get? Filter
vector = indirect-injectionon the site and readauthorityandoutcome. - Is there a dataset of LLM security incidents mapped to the OWASP LLM Top 10, the OWASP Agentic Top 10 and MITRE ATLAS? Every record carries the three mappings;
stats.jsoncounts them. - What has gone wrong with MCP servers, coding agents and AI IDEs in practice? Search the site for a product name, or filter by
affected.productsin the JSON. - How do I cite real AI agent incidents in a threat model, a risk register or a paper? Each record has a primary source, a stable id and a permanent page;
CITATION.cffgives the dataset citation. - How do I get notified of new AI agent incidents? Subscribe to the RSS feed.
Is there a public dataset of AI agent security incidents? Yes, this one: 88 publicly documented events from February 2023 to September 2026 (29 incidents, 48 vulnerability disclosures, 11 threat reports), one schema-validated JSON record each, coded under a written codebook and mapped to the OWASP Top 10 for LLM Applications, the OWASP Top 10 for Agentic Applications and MITRE ATLAS. Download it as JSON or CSV, browse it on the site, subscribe to the RSS feed, or load the Hugging Face mirror. The data is CC BY 4.0. Current counts are in docs/stats.md.
How are incidents coded?
Every event is coded from a primary source that was opened and read, on eight fields defined in docs/codebook.md: type (incident, vulnerability disclosure or threat report), lens (the role AI plays: weapon, target or surface), vector (how the attack or failure got in), channel_in, authority (what the AI component could do), channel_out (how the effect left the system), adversarial (whether an attack technique is involved) and outcome (the most severe harm realised or demonstrated). Each record also carries mappings to the OWASP LLM Top 10, the OWASP Agentic Top 10 and MITRE ATLAS, given only where the source supports them. The paper behind the seed data reports two further blind codings and their agreement (kappa 0.82 for lens and 0.87 for vector with the second coder).
Can I use it commercially? Yes. The data is licensed CC BY 4.0: use, copy, modify and redistribute it, including in commercial products, as long as you credit the dataset (name it and link to this repository) and say if you changed it. The build and validation code is MIT. There is no warranty, and the dataset is a convenience sample of what was made public, so no share computed from it estimates a population share.
How do I add an incident?
One event is one JSON file and one pull request: copy an existing record, give it the next free id, fill every field from a public primary source, run python3 scripts/validate.py, and open the pull request. If you would rather not write JSON, use the issue form. Only events with a public primary source are accepted; this is not a place to disclose new vulnerabilities. The checklist is in CONTRIBUTING.md.
How do I cite the dataset? Use CITATION.cff (GitHub shows it under "Cite this repository"): Muhammad Basit Ali, AI Agent Incidents: an open dataset of publicly documented AI-agent and LLM-application security incidents, version 1.2.0, 2026, https://github.com/basitalisandhu/ai-agent-incidents. The coding scheme comes from the paper AI as Weapon, Target, and Surface: A Threat Taxonomy and a Deterministic Control Plane for Securing LLM Agents (Ali, 2026), whose codebook is reproduced in docs/codebook.md; cite both when you use the coding.
What can the dataset not tell you? It is a convenience sample of events that were made public, so it over-represents what vendors and researchers chose to disclose and says nothing about how common any class of event is in the population. Mappings are the maintainer's reading of each source against the published frameworks; an empty mapping list means no confident mapping, not that none applies. Dates are the month of the public report, not of the event. Each record links its primary source so every claim can be checked.
| Records | 88 events (29 incidents, 48 vulnerability disclosures, 11 threat reports) |
| Period | February 2023 to September 2026 (month of public report) |
| Unit | One primary source's account of one incident, advisory or report |
| Coded fields | type, lens, vector, channel_in, authority, channel_out, adversarial, outcome |
| Identifying fields | id, date, name, cve, sources, summary |
| Added fields | mappings (OWASP LLM Top 10 2025, OWASP Agentic Top 10 2026, MITRE ATLAS), affected (vendors, products, frameworks), tags, status |
| Formats | JSON (one file per event, schema), flat CSV (data/incidents.csv, the paper's original column set), site exports (JSON array, JSONL, CSV, RSS, stats) |
| Coding | By the author of the paper under docs/codebook.md; the paper reports two further blind codings and their agreement |
| Sampling | Convenience sample of what was public; it over-represents what vendors and researchers chose to disclose, and no share computed from it estimates a population share |
| Licence | Data CC BY 4.0, code MIT |
Full definitions, with the number of seed events per value, are in
docs/codebook.md. Current counts are in docs/stats.md, and the OWASP Agentic Top 10
mappings cross-tabulated against the coded fields are in
docs/owasp-agentic-crosstab.md
(python3 scripts/agentic_crosstab.py).
- lens: the role AI plays.
surface(47): the AI is a conduit to another party's assets.weapon(14): the AI is the attacker's instrument.target(19): the AI system itself is attacked, exposed or disrupted. - vector: how the attacker content or the failure got in, for example
indirect-injection(27),autonomous-ops(10),exploitation(7),supply-chain(5),excessive-agency(4),nhi-secrets(4). - channel_in: where the content came from (
chat message,web page,document,email,repo issue/pr,support ticket,calendar invite,tool description,rules file,package,none). - authority: the capability the component or credential gave the event (
none,read-only,send-message,shell/exec,database,write-repo,cloud-creds,file-delete,payments). - channel_out: how the effect left the system (
tool-call send,image/link fetch,code exec,file publish,data destruction,financial transfer,api-abuse,service-disruption,disclosure-only). - adversarial: whether an attack technique is part of the event (65 yes, 15 no).
- outcome: the most severe harm class realised or demonstrated (
data-exfiltration,information-disclosure,code-execution,data-destruction,financial-loss,fraud,service-disruption,none-demo).
Each record carries mappings.owasp_llm, mappings.owasp_agentic and
mappings.mitre_atlas. They are the maintainer's reading of the primary source
against the published frameworks, given only where the source supports them; an empty
list means "no confident mapping", not "none applies". Events with no adversary carry
no ATLAS techniques, because ATLAS describes adversary behaviour. Coverage of the seed
data: 62 events mapped to the OWASP LLM Top 10, 51 to the OWASP Agentic Top 10, 66 to
ATLAS. Mapping suggestions are welcome as pull requests with a one-line justification.
Every seed event was coded from a primary source that was opened and read. On
2026-10-03 the maintainer re-checked the 80 URLs from a network that allows only some
hosts: 12 were reachable and matched their records; 68 could not be reached because the
proxy refused the connection, which says nothing about the pages themselves. Details
per URL are in docs/verification-log.md. No record was
downgraded from confirmed. On 2026-10-13 all 96 source URLs of the 88 records
resolved from an unrestricted network (scripts/check_sources.py, which a weekly
workflow now runs); none was dead.
{
"id": "032",
"date": "2025-06",
"name": "EchoLeak zero-click exfiltration from Microsoft 365 Copilot",
"type": "vulnerability-disclosure",
"lens": "surface",
"vector": "indirect-injection",
"channel_in": "email",
"authority": "read-only",
"channel_out": "image/link fetch",
"adversarial": true,
"outcome": "data-exfiltration",
"cve": ["CVE-2025-32711"],
"sources": [{"url": "https://thehackernews.com/2025/06/zero-click-ai-vulnerability-exposes.html"}],
"summary": "Zero-click via crafted email, CVSS 9.3; fixed server-side; no known exploitation.",
"mappings": {"owasp_agentic": ["ASI01"], "owasp_llm": ["LLM01", "LLM02"], "mitre_atlas": ["AML.T0051.001", "AML.T0077"]},
"affected": {"vendors": ["Microsoft"], "products": ["Microsoft 365 Copilot"], "frameworks": []},
"tags": ["zero-click", "email", "markdown-image-exfiltration"],
"status": "confirmed"
}import json, urllib.request
url = "https://basitalisandhu.github.io/ai-agent-incidents/incidents.json"
events = json.load(urllib.request.urlopen(url))
exfil_by_tool = [e for e in events if e["channel_out"] == "tool-call send"]
print(len(exfil_by_tool), "events exfiltrated through a tool call the attacker influenced")# Events where an agent with shell access was steered by content it read
jq '[.[] | select(.authority == "shell/exec" and .vector == "indirect-injection") | .name]' site/incidents.jsonFor streaming ingestion, site/incidents.jsonl contains one UTF-8 JSON object per
line, in id order, with exactly the same fields as the JSON array. Each build
regenerates both exports; rebuilding unchanged records produces identical bytes.
jq -c 'select(.authority == "shell/exec")' site/incidents.jsonlSubscribe to feed.xml to be notified of new records.
incidents/ one JSON record per event: <id>-<slug>.json (source of truth)
schema/ incident.schema.json (JSON Schema 2020-12)
data/incidents.csv flat export in the paper's fourteen columns (generated)
docs/ codebook.md, stats.md and owasp-agentic-crosstab.md (generated), verification-log.md
scripts/ import_csv.py, export_csv.py, agentic_crosstab.py, digest.py, validate.py, build_site.py, check_sources.py, push_to_hf.py, records.py (the shared safe reader), indexnow_key.txt
digest/ monthly digests, YYYY-MM.md (generated by digest.py, linked from the site)
assets/ og-image.png, copied into the site for link previews
site/ generated static site (index, per-event pages, feed, JSON, CSV, stats, sitemap, robots, llms.txt, IndexNow key)
tests/ pytest suite
hf/ README.md, the dataset card for the Hugging Face mirror
.github/ CI (validate, test, build), Pages deployment with an IndexNow ping, weekly source-link check, Dependabot, issue form
git clone https://github.com/basitalisandhu/ai-agent-incidents
cd ai-agent-incidents
python3 -m venv .venv && . .venv/bin/activate # CI uses Python 3.12
pip install -r requirements-dev.txt
python3 scripts/validate.py # schema and integrity checks
python3 -m pytest -q # the same, plus round-trip and date checks
python3 scripts/build_site.py # regenerates site/ and docs/stats.md
python3 scripts/export_csv.py # regenerates data/incidents.csv from the JSON records
python3 scripts/export_csv.py --since 2025-01 --until 2025-12 --out 2025.csv
python3 scripts/import_csv.py --csv some.csv # imports rows from a CSV in the paper's format
python3 scripts/digest.py --month 2026-10 # writes digest/2026-10.md from the records added that month
python3 scripts/check_sources.py --wayback # requests every source URL; exit 1 if one is dead
HF_TOKEN=... python3 scripts/push_to_hf.py # uploads the CSV, JSON, schema and card to the Hugging Face mirrorCSV month filters are optional and inclusive, accept YYYY-MM, and preserve id
ordering. Without filters, the export remains identical to the full dataset.
Invalid months or a reversed range exit with code 2 without touching the output.
The scripts use only the standard library, except validate.py and the tests, which
need jsonschema. The site has no external dependencies and works when opened from a
local file.
A monthly digest (digest/YYYY-MM.md) summarises the records added to the dataset in
one month: counts by type, lens, vector and outcome, the vendors named more than once,
and a linked list of the records. "Added" is the month a record's file was first
committed, so digest.py needs the full git history (in a shallow clone, run
git fetch --unshallow first). The site's front page links every digest.
| Message | Cause and fix |
|---|---|
incidents/090-x.json: (root): 'summary' is a required property (from validate.py) |
The record is missing a field; every field in the schema is required. Copy an existing record as a starting point. |
error: incidents/090-x.json: missing field 'date'; run python3 scripts/validate.py for every problem |
export_csv.py, build_site.py, agentic_crosstab.py and check_sources.py stop at the first record they cannot use. Run validate.py, which lists every problem in every file. |
error: incidents/090-x.json: not UTF-8 text (byte 0); save it as UTF-8 |
The file was saved as UTF-16 or another encoding, as some Windows editors and PowerShell > do. Save it as UTF-8. |
error: incidents/090-x.json: invalid JSON: Expecting ',' delimiter: line 12 column 5 (char 301) |
A missing comma or quote at that line. python3 -m json.tool incidents/090-x.json shows the same position. |
error: some.csv line 3: expected 14 fields or adversarial must be yes or no (from import_csv.py) |
A short row (often an unquoted comma in notes) or a value outside yes/no. Nothing is written until every row is valid. |
error: the clone is shallow, so first-commit dates are unknown; run git fetch --unshallow |
digest.py needs the history to know when each record was added. |
export_csv.py: error: argument --since: expected a month in YYYY-MM format |
Months are zero-padded, for example 2025-01. |
ModuleNotFoundError: No module named 'jsonschema' |
Activate the venv and run pip install -r requirements-dev.txt. |
One event is one JSON file and one pull request. Copy an existing record, give it the
next free id, fill every field from the primary source, run validate.py, and open the
pull request. CONTRIBUTING.md has the checklist and the source
requirements; the issue form is the way to
propose an event without writing JSON. Only events with a public primary source are
accepted; this is not a place to disclose new vulnerabilities (see
SECURITY.md).
Cite the dataset and the paper it was created for. CITATION.cff is machine-readable (GitHub shows a "Cite this repository" button).
Muhammad Basit Ali. AI Agent Incidents: an open dataset of publicly documented
AI-agent and LLM-application security incidents. Version 1.2.0, 2026.
https://github.com/basitalisandhu/ai-agent-incidents
Muhammad Basit Ali. AI as Weapon, Target, and Surface: A Threat Taxonomy and a
Deterministic Control Plane for Securing LLM Agents. 2026. Manuscript.
- Add archived copies as second sources (issue #6). Every source URL resolved on 2026-10-13, so this is protection against future moves, not a repair.
- Review the OWASP Agentic Top 10 mappings against the published document at genai.owasp.org once it can be fetched; the ids were cross-checked against the OWASP GenAI Security Project's GitHub repositories.
- Add a
controlsfield: the control (provenance, approval, least authority, egress mediation) that would have stopped each event. - Keep adding events after September 2026.
See CONTRIBUTING.md and the open good first issues. The plan is in ROADMAP.md; questions and ideas go in Discussions.
Thank you to everyone who has had a pull request merged here:
- @ruthraguru248 added the vendor statistics table (#13).
- @nawaaaaaAaar added inclusive month filters to the CSV exports (#12).
Data (incidents/, data/, and the site's JSON, CSV and feed exports) is licensed
under CC BY 4.0. Code (scripts/, tests/, workflows, site
templates) is licensed under MIT. Attribution: Muhammad Basit Ali,
https://github.com/basitalisandhu/ai-agent-incidents.
More tools by the same author: https://github.com/basitalisandhu
- agentic-semgrep-rules: Semgrep rule pack for insecure agent code: unbounded tool permissions, eval of model output, SSRF through tool URLs, prompt interpolation, MCP servers without auth.
- agent-threat-model: CLI that turns a YAML description of an agent system into a STRIDE + OWASP Agentic threat model, control checklist and Mermaid diagram.
- agent-security-skills: Claude Code plugin and agentskills-compatible skill pack for agent security reviews: threat modelling, config audits, policy generation, incident lookup.