ANY MODEL. THE SAME EVIDENCE.

Use OpenAttacks with your LLM

The agent skill is the path for Claude, Cursor, and other coding agents. It reads the slim index, then fetches only the attack files it needs. A chat window should paste one category or one month, not the whole catalog. This page is built from 275 current entries.

Install the openattacks skill

The skill is skills/openattacks/SKILL.md. Its name is openattacks, matching the folder. The zip contains openattacks/SKILL.md and nothing else. After it is installed, ask a question about a reported AI attack. The agent loads the skill when the question matches that description, or you can call it as /openattacks where the product uses slash commands.

Download openattacks-skill.zip

Each command downloads that zip and unpacks it so SKILL.md lands in the right folder. Run a project command from the repository root when the team should share the skill. Run the home-directory command when you want it in every project on your machine. On Windows, the same zip unpacks with PowerShell. This one installs it for Claude Code; change the destination for another tool:

Invoke-WebRequest https://openattacks.ai/downloads/openattacks-skill.zip -OutFile $env:TEMP\openattacks-skill.zip; New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.claude\skills" | Out-Null; Expand-Archive -Force $env:TEMP\openattacks-skill.zip "$env:USERPROFILE\.claude\skills"

Claude Code

Project skills live at .claude/skills/openattacks/SKILL.md. Personal skills live at ~/.claude/skills/openattacks/SKILL.md and apply to every project on that machine. Extend Claude with skills is the directory list and the frontmatter reference.

This project

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p .claude/skills && unzip -o /tmp/openattacks-skill.zip -d .claude/skills

Every project on this machine

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p ~/.claude/skills && unzip -o /tmp/openattacks-skill.zip -d ~/.claude/skills

Claude.ai and Claude Desktop

Skills need code execution. Turn on Code execution and file creation under Settings > Capabilities. On Team and Enterprise plans an Owner enables it under organization settings. Then open Customize > Skills, choose +, and upload the zip. Use skills in Claude documents that flow. The zip must contain the openattacks folder, not a bare SKILL.md at the zip root (how to package a skill). Turn the skill on. In a chat, type / and select it, or ask a question the description matches.

Claude Desktop uses the same Customize > Skills page. A skill you upload there stays on that device (Skills overview,Skills in Claude Desktop). Uploading it to claude.ai does not install it in Claude Code, and the reverse is also true.

Cursor

Put the folder at .cursor/skills/openattacks/SKILL.md for one project, or ~/.cursor/skills/openattacks/SKILL.md for every project on the machine. Cursor also reads .agents/skills/, .claude/skills/, and .codex/skills/. Agent Skills describes discovery and the /openattacks command. Rules are a different mechanism: they load on every turn (Rules). Use the skill for this catalog. Use a rule only if you want a one-line pointer in every chat.

This project

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p .cursor/skills && unzip -o /tmp/openattacks-skill.zip -d .cursor/skills

Every project on this machine

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p ~/.cursor/skills && unzip -o /tmp/openattacks-skill.zip -d ~/.cursor/skills

OpenAI Codex CLI and ChatGPT desktop

Install new skills at ~/.agents/skills/openattacks/SKILL.md, or .agents/skills/openattacks/SKILL.md inside a repo. Build skills is the current location list. The same guide covers ChatGPT desktop, where you can mention a skill with @.~/.codex/skills/ is the old path. Codex still reads it, and new installs should use ~/.agents/skills/.

AGENTS.md is not a skill. Codex adds it to every turn, and the project file is capped (Custom instructions with AGENTS.md). Keep the skill for the catalog procedure. If you want a standing reminder, put the short instructions from the chat section into AGENTS.md.

This project

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p .agents/skills && unzip -o /tmp/openattacks-skill.zip -d .agents/skills

Every project on this machine

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p ~/.agents/skills && unzip -o /tmp/openattacks-skill.zip -d ~/.agents/skills

GitHub Copilot

Copilot loads agent skills from .github/skills/openattacks/SKILL.md in a repo, and from ~/.copilot/skills/ or ~/.agents/skills/ for you personally. It also reads .claude/skills/ and .agents/skills/. About agent skills and Adding agent skills describe when Copilot loads the file.

.github/copilot-instructions.md is added to every request in that repository (repository custom instructions). Use it for a short pointer. Keep the procedure in the skill.

This project

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p .github/skills && unzip -o /tmp/openattacks-skill.zip -d .github/skills

Every project on this machine

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p ~/.copilot/skills && unzip -o /tmp/openattacks-skill.zip -d ~/.copilot/skills

Gemini CLI

Workspace skills go in .gemini/skills/openattacks/SKILL.md. User skills go in ~/.gemini/skills/openattacks/SKILL.md..agents/skills/ is an alias and takes precedence when both exist. Agent Skills documents the tiers. In the CLI, /skills list should show openattacks. Workspace skills load only in a trusted folder.

GEMINI.md is sent on every prompt (GEMINI.md files). Put a one-line pointer there if you want it always, and leave the procedure in the skill.

This project

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p .gemini/skills && unzip -o /tmp/openattacks-skill.zip -d .gemini/skills

Every project on this machine

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p ~/.gemini/skills && unzip -o /tmp/openattacks-skill.zip -d ~/.gemini/skills

Windsurf

Cascade reads .windsurf/skills/openattacks/SKILL.md in the project and ~/.codeium/windsurf/skills/openattacks/SKILL.md globally. It also discovers .agents/skills/. Mention @openattacks or ask a question that matches the description. The directory list is in the Cascade skills article (the English URL currently redirects away from that article) and in the Devin CLI skills overview, which lists .windsurf/skills/ as well.

This project

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p .windsurf/skills && unzip -o /tmp/openattacks-skill.zip -d .windsurf/skills

Every project on this machine

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p ~/.codeium/windsurf/skills && unzip -o /tmp/openattacks-skill.zip -d ~/.codeium/windsurf/skills

Cline and Roo

Cline reads .cline/skills/openattacks/SKILL.md and ~/.cline/skills/openattacks/SKILL.md, and also .claude/skills/(Cline skills). The name in the file must match the folder name.

Cline, this project

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p .cline/skills && unzip -o /tmp/openattacks-skill.zip -d .cline/skills

Cline, every project on this machine

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p ~/.cline/skills && unzip -o /tmp/openattacks-skill.zip -d ~/.cline/skills

Roo reads .roo/skills/openattacks/SKILL.md and ~/.roo/skills/openattacks/SKILL.md. Project skills override a global skill of the same name (Roo skills). Roo rules in .roo/rules/ are always on (custom instructions); prefer the skill.

Roo, this project

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p .roo/skills && unzip -o /tmp/openattacks-skill.zip -d .roo/skills

Roo, every project on this machine

curl -fsSL https://openattacks.ai/downloads/openattacks-skill.zip -o /tmp/openattacks-skill.zip && mkdir -p ~/.roo/skills && unzip -o /tmp/openattacks-skill.zip -d ~/.roo/skills

ChatGPT and Gemini on the web

These chats do not read a skills folder on your computer. Give them the procedure, then one bundle from the next section.

Instructions to paste

Read https://openattacks.ai/llms.txt first. It is a slim index. Then open only the entries this question needs at https://openattacks.ai/llms/attacks/<id>.md. For a whole category or month, open https://openattacks.ai/llms/by-category/<category>.md or https://openattacks.ai/llms/by-month/<YYYY-MM>.md instead. For attacks that used one offensive framework, open https://openattacks.ai/llms/by-framework/<id>.md. Do not load llms-full.txt. Cite the catalog page and the original source URL. Say when attribution or AI involvement is unknown. Do not invent detection rules or MITRE ATLAS mappings.
  • ChatGPT Projects.Create a project, open the more-options menu, and add the instructions above under Project settings. Upload one category or month file as project knowledge. Projects in ChatGPT. Custom instructions under Settings > Personalization apply to every chat and are short (custom instructions), so a project is the better fit.
  • Custom GPTs.If you still have one, paste the same instructions into the GPT instructions field (creating GPTs). OpenAI is retiring custom GPTs; new creation has already ended for some workspaces, with standard retirement scheduled for December 11, 2026. Prefer a project.
  • Gemini Apps skills.On gemini.google.com open Customize > Skills, then Create > Upload. Upload the unzipped openattacks folder so SKILL.md sits in that folder, or upload SKILL.md by itself. Create and manage skills for Gemini Apps. Personal accounts are moving Gems to skills (about the transition). Where Gems remain, paste the instructions into a new Gem and attach one bundle under Knowledge (Use Gems in Gemini Apps).

Any other agent that accepts a system prompt or a SKILL.md file can use the same folder. The format is the Agent Skills specification: a directory, YAML name and description, then markdown instructions.

Paste a bundle, or point at the index

llms-full.txt and the full JSON catalog are the complete record. They are the wrong first paste for a chat: they crowd out the question. Paste one bundle. The names below are the categories and months in this build. A web-enabled model can fetch the index itself and open only the files it needs.

Category bundles

16 entries have no category. They stay in the index and in their month bundle.

Offensive framework bundles

A bundle exists only when an attack's frameworks field names that tool. A passing mention does not.

Recent month bundles

Older months use the same path, llms/by-month/YYYY-MM.md. One attack is llms/attacks/<id>.md.

Structured files

Upload api/attacks.json when the model can run code and you want every field. In a git checkout the same records are data/attacks.json. api/schema.json is the field list. llms-full.txt remains available for a tool that can hold the entire catalog.

Web-enabled models

Give the model this URL and tell it to open the index before it opens any attack. ChatGPT with browsing, Claude with web access, and other tools that can fetch a URL can follow the links inside it.

https://openattacks.ai/llms.txt

A question that works with that URL: “Open https://openattacks.ai/llms.txt. Which attacks in the latest month targeted finance, and how confident is the attribution? Fetch only those attack files.”

Example prompts

These prompts name entries from the current catalog, so the examples move when the YAML changes. They work with the skill, with a pasted bundle, or with the index URL.

Threat intelligence

Financial-sector attacks

Which AI attacks targeted banks or other financial institutions in the last 30 days? For each match, say who attributed it, the attribution confidence, and how confident the AI involvement is. Include “SOCRadar XaocMail ClickFix delivers Phantom53 and HavoX, built with Claude Code” (socradar-xaocmail-phantom53-claude-code-2026-11) and “Island TypoPrompting turns AI-brand typos into ClickFix, Net Guard, and search hijacks” (island-typoprompting-ai-typosquats-2026-11), then every other finance-sector entry the index shows. Cite the entry id and source URLs.

Who used a named tool

List every catalog entry that used ARTEX. For each, give the report date, targeted region or sector, attribution and its confidence, and the AI involvement confidence. Start with “Eye Security ARTEX intrusion from a PHP webshell to GitOps and CI runners” (eye-security-artex-gitops-2026-11); “Arctic Wolf - AntiFox ARTEX fork web-shells IIS and reads AWS production secrets” (arctic-wolf-antifox-iis-aws-2026-11); “Railway October 2026 incident - build escape to decrypted customer AI provider keys” (railway-october-ai-key-theft-2026-10). Do not merge separate reports into one actor.

Detection engineering

Rules for the ARTEX intrusions

Give me the Sigma, Splunk, and Elastic rules recorded for the ARTEX intrusions, including “Eye Security ARTEX intrusion from a PHP webshell to GitOps and CI runners” (eye-security-artex-gitops-2026-11); “Arctic Wolf - AntiFox ARTEX fork web-shells IIS and reads AWS production secrets” (arctic-wolf-antifox-iis-aws-2026-11); “Railway October 2026 incident - build escape to decrypted customer AI provider keys” (railway-october-ai-key-theft-2026-10). For each rule, include the format, rule name, URL when one is present, and the notes. If a format has no rule on that entry, say so. Do not invent rules that are not in the entry.

Detection gaps

For these ARTEX entries — “Eye Security ARTEX intrusion from a PHP webshell to GitOps and CI runners” (eye-security-artex-gitops-2026-11); “Arctic Wolf - AntiFox ARTEX fork web-shells IIS and reads AWS production secrets” (arctic-wolf-antifox-iis-aws-2026-11); “Railway October 2026 incident - build escape to decrypted customer AI provider keys” (railway-october-ai-key-theft-2026-10) — which only have guidance detections, and which have a Sigma, Splunk, Elastic, or KQL rule? Summarize each guidance note in one line and name the entry id it came from. Do not add rules that are not in the entry.

MITRE ATLAS mapping

Map recent entries

Map the attacks reported in 2026-11 to MITRE ATLAS. For each entry that has atlas_techniques, list the technique id, the rationale, and the source URL. If an entry has no mapping, say it is unassessed. Do not treat a missing mapping as evidence the technique did not occur.

One intrusion, one mapping

Take “Eye Security ARTEX intrusion from a PHP webshell to GitOps and CI runners” (eye-security-artex-gitops-2026-11). List each MITRE ATLAS technique id on that entry, why it was mapped, and the source URL that supports the rationale. Keep the mapping separate from the incident facts.

Prevention and hardening

Harden against ARTEX use

From the prevention_angle fields, what should a defender harden first for the ARTEX cases “Eye Security ARTEX intrusion from a PHP webshell to GitOps and CI runners” (eye-security-artex-gitops-2026-11); “Arctic Wolf - AntiFox ARTEX fork web-shells IIS and reads AWS production secrets” (arctic-wolf-antifox-iis-aws-2026-11); “Railway October 2026 incident - build escape to decrypted customer AI provider keys” (railway-october-ai-key-theft-2026-10)? Group the controls, keep each one tied to the entry id it came from, and do not present them as controls that were verified during the incident.

Agent and MCP target surface

Which entries have ai_role ai_agent_or_mcp_as_target? Summarize the initial access and the prevention_angle for the five most recently reported. Call out anything a team should verify in their own agent, IDE, or MCP client configuration.

Executive briefing

One-month briefing

Write a one-page executive briefing on the AI attacks reported in 2026-11. Use only that month's entries. Cover who was targeted, whether AI was the tool or the target, what is known about attribution and how confident it is, and three defensive priorities. Keep it non-technical and cite entry names.

What not to overclaim

For a leadership update on “SOCRadar XaocMail ClickFix delivers Phantom53 and HavoX, built with Claude Code” (socradar-xaocmail-phantom53-claude-code-2026-11) and “Island TypoPrompting turns AI-brand typos into ClickFix, Net Guard, and search hijacks” (island-typoprompting-ai-typosquats-2026-11), separate confirmed AI involvement from vendor-asserted or unverified involvement, and separate high-confidence attribution from medium, low, or missing attribution. State the unknowns plainly.

Attacks on AI vs adversary use of AI

Compare the two categories

Compare attacks on AI with adversary use of AI. Use the category field: attacks-on-ai versus adversary-use-of-ai. Give the count of each, two examples of attacks on AI (start with “Dynatrace mcpvenom measures MCP tool-description poisoning across 12 frameworks” (dynatrace-mcpvenom-tool-poisoning-2026-11) and “Elastic REF3364 SANSAR takes over exposed Ray GPU hosts and hides utilization” (elastic-sansar-ray-gpu-2026-11)), and two examples of adversary use (start with “Elastic CyberStrikeAI server reads AWS user data, secrets, and sends an SSM command” (elastic-cyberstrikeai-aws-cloudtrail-2026-11) and “SOCRadar XaocMail ClickFix delivers Phantom53 and HavoX, built with Claude Code” (socradar-xaocmail-phantom53-claude-code-2026-11)). Explain how the prevention notes differ. Leave uncategorized entries out of both counts.

Same tool, different side

ARTEX appears both as an attacker tool and as a backdoored tool. Using “Eye Security ARTEX intrusion from a PHP webshell to GitOps and CI runners” (eye-security-artex-gitops-2026-11); “Arctic Wolf - AntiFox ARTEX fork web-shells IIS and reads AWS production secrets” (arctic-wolf-antifox-iis-aws-2026-11); “Railway October 2026 incident - build escape to decrypted customer AI provider keys” (railway-october-ai-key-theft-2026-10), say which entries are adversary-use-of-ai and which are attacks-on-ai, and what a defender would do differently in each case.

Fields the skill expects

Every entry has an id, name, report date, sources, and a prevention note. AI involvement is one primary role, plus an optional secondary role:

  • ai_agent_or_mcp_as_target — Agents, IDEs and MCP connections become the attack surface.
  • attacker_uses_ai_as_operator — Attackers use models and agent loops to carry out intrusion steps.
  • ai_tool_as_lure — AI branding persuades victims to install malware or share access.
  • llm_generated_malware — Models generate, rewrite or obfuscate malicious code.
  • ai_credential_theft — Stolen keys and sessions provide access to AI services.
  • other — Deepfake fraud and other documented AI-enabled techniques.
  • ai_infra_as_c2 — AI services carry commands or support an implant control channel.

Attribution confidence and AI-involvement confidence are separate. A missing value is unassessed. Detection formats are sigma, splunk, elastic, kql, lolbas, and guidance. The methodology explains sourcing, geography, and what the catalog refuses to infer.