Blacklight
Agentic defense layer for the open-source Linux stack
LandingVisit the Blacklight landing pageBlacklight turns the open-source Linux defensive stack into a system that can investigate and respond to incidents on its own. It uses Linux Malware Detect as the post-scan trigger, ModSecurity and Apache logs as evidence, and iptables as the response surface. No new agents on the host. No EDR contract. A shell script and an API key.
The system is built around three primitives that only became viable in 2026: Opus 4.7 with a million-token context window, Anthropic Managed Agents for a curator session that survives across operators and hosts, and the skills-native pattern for description-routed lazy-loaded behavior. The model choice is the system design.
Blacklight is model-bounded today. It operates at the edge of what's possible across roughly a five-system incident, which covers over 80% of real-world cases. It works now, and it improves as the limits move. The dedicated landing page at /blacklight is the canonical home: architecture, walkthrough, skills, roadmap, build timeline, and the closing love letter to operators.
Features
Detection & Response
- Post-scan hook fires from Linux Malware Detect into a Blacklight forensic case
- Read-only kill-chain extraction: Apache transfer log + matching ModSecurity audit transactions
- Curator emits a defense action into the memstore with full reasoning trace
- ModSec rule lands behind an atomic symlink with Apache config test passing
- Known bad actor IPs go straight into the live iptables chain on the running firewall
Case Persistence
- Anthropic Managed Agents curator session, resumable across a 30-day checkpoint window
- The case persists across operators and hosts; the host doesn't
- Three days later, different operator, different host: reattach and the ledger is right there
- Description-routed, lazy-loaded skills (six routing skills, eight workspace corpora)
Substrate & Model
- Opus 4.7 with a million-token context window holds the whole investigation in-window
- No retriever drops the one record that mattered
- Skills-native pattern: twenty-three skill directories, operator-voice behavior
- Weaponizes the open-source stack, doesn't replace it: maldet, ModSecurity, Apache, iptables
Honest Limits
- Model-bounded today, even on Opus 4.7
- Operates at the edge of what's possible across roughly a five-system incident
- That window covers over 80% of real-world cases
- Improves as the model context and reasoning limits move
Installation
$ git clone https://github.com/rfxn/blacklight.git
$ cd blacklight
$ sudo ./install.sh