ThreatDefence/agentic-attack-detections
Security & PrivacyDetection rules and indicators for intrusion frameworks that drive AI agents to run attacks.
Overview
Agentic attack frameworks hand an intrusion to AI agents that plan, execute real attack tooling and persist until a goal is reached. This repository packages detection content for that class of tooling: KQL hunts, Microsoft Sentinel analytics rules, Sigma rules, Suricata signatures, Zeek scripts and intel files, YARA rules, plus shared indicators and machine-readable feeds.
It covers two layers. Per-framework folders target specific tools — ARTEX (and its AntiFox rename), Cairn, PentAGI, HexStrike, CyberStrikeAI, Strix, Hermes Agent and Agent Zero — each built from a review of the framework's source code and public reporting. The behaviors/ folder holds rules that are not tied to one framework, such as agent CLIs running attack tools in auto-approve mode, servers calling model APIs, rotating-proxy probing, and callbacks to the probing address.
The material is aimed at detection engineers and SOC analysts running Sentinel, Defender XDR, Suricata, Zeek, YARA or Splunk/Elastic, not at users deploying attack frameworks themselves.
Getting started
Choose the backend you already operate. For Microsoft Sentinel, upload feeds/sentinel-watchlist.csv as a watchlist with alias AGENTIC_IOCs and search key value, import the analytic rules from frameworks/*/sentinel/analytics/ and behaviors/sentinel/analytics/, and let each framework's watchlist sweep filter the shared list to its own rows.
For network detection, add the .rules files under frameworks/*/suricata/ and behaviors/suricata/ to Suricata; HTTP rules require cleartext or decrypted traffic, while DNS and TLS SNI rules work on encrypted traffic. For Zeek, load the behavior package in behaviors/zeek/agentic-behaviors/ (see the Zeek setup notes), the ARTEX script, and the per-framework .intel files. Use frameworks/*/yara/ to scan container layers, home directories, /tmp and web roots.
Sigma rules can be converted with sigma convert or taken from the pre-converted generated rules. Threat intelligence platforms can import the STIX 2.1 bundle or the per-framework plain lists. Indicators are maintained in iocs/indicators.csv, which is the single source every feed is generated from.
To validate rule changes locally:
pip install -r tools/requirements.txt
python3 tools/validate.py
python3 tools/convert.py
The first command runs checks across Sigma, KQL, ES|QL, TOML, Splunk, Suricata SIDs, YARA, osquery and feeds; the second regenerates the generated/ conversions. Network tests need zeek and suricata on PATH.
Constraints and confidence
The repository states that no rule has been tested against a captured intrusion, so rules should be treated as experimental until they run against your own data. Each indicator carries a confidence level: high for framework-specific user agents, command lines, paths and hashes; medium for reported addresses and shareable prompt text, suitable for alerting after baselining; low for common items such as model APIs, request-collection services and proxy providers, to be used only as context.
Addresses from cloud and VPS providers get reassigned; each has a review_by date 180 days after reporting and should be checked before blocking. Suricata HTTP rules need decrypted traffic. A rule match is a lead, not a verdict, and no match does not prove absence of compromise.
Extending the content
New indicators go into iocs/indicators.csv with their source and framework, then python3 tools/build_feeds.py and python3 tools/validate.py regenerate feeds and check them. Adding a whole framework follows the framework addition guide; false positives are reported as issues with the rule name and event shape. The project is MIT licensed and indicators can be reused in other products and feeds.
HotRepo's take
The repository is very new, and the only recorded activity is a new-project Python list entry on 2026-11-22, where it gained 8 stars in one hour. There is no GitHub daily-rank history. What stands out technically is the breadth of backends covered from a single indicator CSV — Sigma, Splunk SPL, Sentinel ASIM KQL, Elastic queries, Suricata, Zeek and YARA — which lowers the cost of adopting the same detections across whatever stack a SOC already runs.
The explicit caveat that no rule has been tested against a captured intrusion is the key fact for evaluation: value depends on testing against local telemetry and tuning false positives before alerting.
Written by AI from the project's public information and HotRepo's records, updated . The project's own docs are the source of truth.
You might also like
- SigmaHQ/sigmaA community-maintained library of vendor-neutral Sigma detection rules for SIEM and log analysis, aimed at detection engineers and threat hunters.
- ikow/php2shell-pocExploit and detection toolkit for two critical PHP pre-auth RCE bugs affecting nearly every default PHP install.
- ikow/java4shell-pocPoC, scanners and detection rules for an OpenJDK AVX-512 Base64 heap overflow affecting JDK 17–27.
- failsafe-security/artexBuilds a locally bound, source-compiled ARTEX security testing harness from audited source.