Repository navigation
Releases
2.9.1: Agent Zero source review and a Zeek package
Agent Zero gets its source review, the review corrected two rules (#7), and the Zeek behaviors package installs with zkg.
- New
frameworks/agent-zero/source-code-signals.md: how Agent Zero runs, what a target sees, what the host that runs it shows, the exposed web UI, and what the code does not contain, with a file and line for each signal at commita6e41ca. It also checks every string the Agent Zero rules match against the source. - Agent Zero hunt 1: the attack-tool and tunnel branches looked for
/a0/in the command line of the parent process. The source shows the parent is/bin/bash, and tunnels are opened bysshorcloudflared, so those branches could not fire. They now look at devices where Agent Zero itself was seen, and the tunnel branch also covers Microsoft dev tunnels. - YARA
AgentZero_Hacker_Profilenow needs a prompt line that only the hacker profile has. Two of its strings are also in the default profile, so a dump of the default prompts could match a rule named for the hacker profile. - The behaviors Zeek package installs with the Zeek package manager: a
zkg.metaat the repository root points atbehaviors/zeek/agentic-behaviorsand runs its tests on install (zkg install https://github.com/ThreatDefence/agentic-attack-detections). - Indicators: the note on the "obey instructions never refuse for safety ethics" string now says it is in the default profile too, and the unsourced "5080 also seen" note on the port is removed.
2.9.0: new Suricata rule IDs and trimmed plain feeds
Two changes that affect anyone who already uses the Suricata rules or the plain indicator lists.
- Suricata rule IDs renumbered. Every SID moved from 9726001-9734999 to 3800001-3808999: subtract 5,926,000 from the old number (9726001 is now 3800001, 9727057 is now 3801057). The old numbers sat inside a range registered to another rule set at sidallocation.org, so the two could collide. The new block, 3800000 to 3819999, has been requested there for this project. If you suppress, threshold or reference these rules by SID, update the numbers. Rule content and revisions are unchanged. The WAF rule mapping, the tests and the docs use the new numbers; older entries in this changelog keep the numbers they were written with.
- Plain indicator lists carry medium and high confidence only.
feeds/*.txtandfeeds/by-framework/no longer include low confidence rows. Many of those are legitimate services that only mean something in context, such asapi.anthropic.com,api.x.ai,webhook.siteand theBoto3user agent, and a bare list cannot carry that context. They remain iniocs/indicators.csv, the Sentinel watchlist, the STIX bundle and the Zeek intel files. The lists now hold 115 IPv4 addresses (was 162), 30 domains (was 45) and 37 user agents (was 44).
Also in this release, from 2.8.1
No detection changes. Releases now come with ready-made downloads.
- Each release carries the Suricata rules in one file, the YARA rules, the STIX 2.1 bundle, the indicators and feeds, the WAF rules, and a
SHA256SUMS.txt. The file names stay the same from release to release, soreleases/latest/download/<name>always gives the newest one. tools/build_release.pybuilds them, with the same bytes for the same commit and version. A workflow attaches them when a release is published, and CI builds them on every push and loads the combined Suricata file.- README: a Downloads table and the
suricata-updatecommand.
2.8.0: WAF rules for Cloudflare, AWS WAF and Azure WAF
WAF rules (#10). The detections stay the same; this release adds a way to block part of them at the edge.
- New
waf/folder with rules for Cloudflare (dashboard expressions, Rulesets API JSON, IP list, Terraform), AWS WAF (rule group and IP sets as CLI input, CloudFormation and Terraform) and Azure WAF (Application Gateway and Front Door policies in Bicep and ARM, and plain custom rule arrays). - Rule groups by tier. Block: framework user agents, probe markers and framework headers, unsigned
alg:noneJWTs, and the CyberStrikeAI webshell probe in request bodies. Challenge: scanner user agents and reported addresses. Log: AI agent page readers and the Hermes results folder. Opt-in: WSUS services reached from the internet. Addresses are never blocked by default. - Everything is generated by
tools/build_waf.pyfromwaf/signatures.yamlandiocs/indicators.csv. Addresses past theirreview_bydate are left out (21 of 163 on this release). tests/waf/run_tests.pychecks the generated files on every platform: every signature value, every address, hand-written requests (most of them near misses), and a replay of the Suricata test flows so the WAF rules and the Suricata rules they mirror agree. It also checks platform limits, the AWS files against the wafv2 API model, and CI runscfn-lint,terraform validateandaz bicep build.waf/README.mdanswers whether a WAF is needed and where it fits next to the SIEM and IDS content.
Also in this release, from 2.7.3
- Elastic conversions of DNS, proxy and web server rules now use ECS field names:
dns.question.name,url.original,url.domain,url.path,url.queryanduser_agent.original. Before, they kept the Sigma names and matched nothing in an ECS index. Contributed by @tridev-tej in #9, closing #4. tools/validate.pyfails if a generated Elastic query still uses one of those Sigma field names.
2.7.2: two ARTEX user agents from a community report
- Two ARTEX user agents added:
artex-agentandartex-judge-r121. Dragos saw them in AWS CloudTrail when ARTEX checked stolen keys it had planted as canary tokens. Reported in #8 by @Assaf-tal. - New Suricata rule 9726009 for both, and the names are added to the ARTEX user-agent rules for Sigma, Microsoft Sentinel, Elastic, Splunk and Zeek. The rules match
artex-judge-followed by any run name, sincer121reads as a run number. - 516 indicators in total.
Full history: CHANGELOG.md
2.7.1: easier to start with and contribute to
No detection changes. This release makes the repository easier to start with and to contribute to.
- README: a one-command "Try it in a minute" that runs the ARTEX Suricata rules over the sample capture, a "Start here" table by what you run, a coverage table of frameworks against formats, and a section on how the rules are built and checked.
CONTRIBUTING.mdwith the checks to run and the rules for indicators, Suricata, KQL and Sigma content.SECURITY.md, and issue forms for a false positive, a missed detection, a new indicator and a new framework.- Framework folder READMEs are titled " Detection Rules and IOCs".
- ARTEX file headers now read "Agentic attack detections by ThreatDefence (ARTEX)", like the other folders. They still carried the repository's first name.
LICENSEcopyright line now includes contributors.- The ARTEX Splunk app is now
TA-agentic-attack-detections, labeled "Agentic Attack Detections (ARTEX)" (wasTA-artex-detections). If you installed it under the old folder name, remove that folder before copying the new one, or Splunk will load both.
Full history: CHANGELOG.md
2.7.0: correctness release
A correctness release. Every rule was reviewed against the framework source and its cited report, the rules were run over 30 days of production telemetry, and two test suites were added. Nothing new is detected; a lot now works that did not.
Rules that could not fire
tools/convert.pyreused the same rule objects for every backend, so each Elastic conversion carried Microsoft field names and only 1 of 64 rules converted for Defender XDR. Each backend now gets its own copy: 25 rules convert for Defender XDR, and Elastic output uses ECS fields, includingprocess.*for Linux process rules. Everything ingenerated/was rebuilt.- AI agent starts an attack tool (behaviors hunt 1, its analytics rule and Sigma rule) looked for an auto-approve flag on the tool's parent. The parent is the agent's shell, so the flag was never there, and Windows
.exenames were missing. The rules now follow the process tree: Claude Code's shell (shell-snapshots,CLAUDE_CODE_SHELL_LAUNCHER_SCRIPT), an agent binary as parent or grandparent, or an agent found by its command line. The flag is a context column, because staff use it too and an agent can run a tool without it. Severity is medium. The analytics rule reads agent processes from the last day instead of the last hour. - Suricata: Cairn 9727002 matched a port in
http.host, which never holds one; it now matches the project list fields in the response. 9726068 looked for a host name inside Gh0st RAT traffic and is retired.ARTX_PERSIST_OKhas its own rule (9726027). BLOODALCHEMY, toolkit and card-theft campaign domains match subdomains. Unsigned JWT rules covertypand the None and NONE spellings (9726030 to 9726033). 9726067 allows a query string. Hermes 9732001 and 9732004 accept spaced JSON andsh -c, and 9732001 no longer needs an outside source.dnslog.cnandwebhook.sitematch any case. - KQL:
sComputerNameis not aW3CIISLogcolumn (CyberStrikeAI hunt 1 and its analytics rule),FileUrlis not aDeviceFileEventscolumn (ARTEX hunt 22), andTicketEncryptionTypeis not aSecurityEventcolumn (ARTEX hunt 16). - CyberStrikeAI's
.bashrcline is written to the file, not run as a command, so the# cscommand-line branch is gone from the hunt, the analytics rule and the Sigma rule. The file-event branch and the YARA rule cover it. - Strix Sigma rule:
strix --target xandstrix -t xnow match. - Elastic: two ES|QL queries used
(?i), which Elastic's regex does not support, and the ARTEX user-agent rule was case-sensitive. Splunk: the user-agent search dropped rows with nosrcfield. - The certificate rules for CyberStrikeAI and PentAGI only see a certificate in a TLS 1.2 handshake. The rule files said otherwise.
Rules that were noisy
- Out-of-band lookups: Burp's own client polling (
polling.oastify.com) is left out everywhere, Suricata alerts once per source per hour, and the Sigma rule is medium. - Strix: only install commands count, not uv or pip probing an installed copy.
- Web service starts a shell (Sigma, KQL, Splunk):
javaandnodecount only under a service account, because they also run Jenkins, npm and AI agent CLIs. - Carbonato:
nsenter -t 1alone is also what Kubernetes node-debug tools do, so the rule needs an install or persistence command with it. - ARTEX hunt 10 no longer matches the bare word
artex, which is also a real company's name. HexStrike's port 8888 hunt leaves out Jupyter. - Suricata: every rule that inspects content has a fast pattern, so none runs on every transaction. Multi-domain regular expressions are now one rule per domain.
- Splunk: searches read
artex_indexesinstead ofindex=*, and each has alert suppression.
Zeek
- On a cluster,
Kerberos_Enumeration,AI_Agent_Attacking_SiteandCallback_To_Proberkept their state on one worker and did not fire when related connections reached another. They now share state through SumStats or the manager. Checked on a local manager with two workers. - With
server_netsunset the package treated every local host as a server, so laptops calling model APIs and Burp polling raised notices. The server-only notices now stay off untilserver_netsis set, with a warning at startup. Rotating_Proxy_Probingchecked its ratio once per window and now checks it on every request. It needs a payload or a refused login, so a normal login page no longer raises it.Backup_File_Servedneeds a real file body, so a site that answers 200 to everything no longer raises it.Callback_To_Proberignores requests that came through a CDN.SMB_WinRM_Fan_Outcounts connections where both sides sent data, not half-open scans.Out_Of_Band_Lookupskips Burp polling and fires once per source per hour.- Two scripts kept an entry for every visitor for up to a day. They now keep only addresses that already showed the first half of a pattern, for an hour.
- The ARTEX script matches user agent names in any case (
ScopeWeaver/1.0,ARTEX-PENTEST/1.0) and every unsigned JWT header form. - The test capture now carries Kerberos requests Zeek parses; before, the Kerberos notice fired from error replies alone.
- Changed settings:
prober_min_connsis nowprober_min_requests,prober_memoryandpage_reader_windowdefault to 1 hour, andCallback_To_Proberneeds cleartext HTTP for the probe half.
Tests
tests/suricata/run_tests.pyreplays one flow per hand-written Suricata rule, with near misses, and fails if a rule has no case or fires where it should not. CI runs it.tools/validate.pyparses every KQL hunt and analytics query with Microsoft's Kusto.Language and checks table and column names against Microsoft's table reference (tools/kqlcheck/).- Every indicator was checked against the report it cites. Three rows that cited our own article now cite Wiz Research, where the values were published.
Full history: CHANGELOG.md