AI-OSINT (LLM-automated intelligence gathering)

AI-OSINT is the use of LLMs and agents to automate open-source intelligence — replacing the analyst who manually runs recon tools and correlates results with a system that queries the web, reads, reasons, and assembles the answer itself. It is the spoke's leading edge: two of its three founding sources are AI-OSINT tools, and they mark a clear progression in autonomy.

A spectrum of autonomy

• Single-LLM loop — llm-osint: one model is given a target, searches the web, and builds a profile. The minimal proof that "LLM + tools → autonomous recon" works.

• Multi-agent orchestration — kallisto-osinter: a search agent, a knowledge-orchestration agent, and specialized deep-dive agents, with LLM map-reduce summarizing large pages. Recursive depth-seeking and staged prompts (its author argues) cut token cost and improve thoroughness over a monolithic call.

• Orchestrated agents that act — strix: the same multi-agent shape, but the specialists divide a kill chain (recon / exploitation / post-exploitation) rather than a reading list, and they run exploits against the target instead of reporting on it. Recon becomes one agent's subroutine.

So the axis runs monolithic LLM → orchestrated agents → agents that act on the target, the same move agent frameworks make generally, pointed first at reconnaissance and then past it (autonomous-pentesting).

What it automates, and why that's a step change

Manual OSINT is gated by analyst time — someone has to run Maltego/theHarvester/Shodan, read the output, and decide what to chase next. AI-OSINT removes that bottleneck, making profile-building cheap, fast, and scalable. That is the capability and the hazard: the same automation that helps an authorized investigator also lowers the cost of mass surveillance and of the reconnaissance behind social-engineering attacks.

How well does the gathering half work? (added 2026-08-22)

The claim above — that automation removes the analyst bottleneck — has been unmeasured on this wiki since spin-out. llm-osint-reliability-study and llm-geolocation-test both score models given evidence. vibesearchbench is the first source here that scores what an agent brought back: expert-built ground-truth graphs, triple-level precision and recall, and a best frontier-model F1 of 30.30, with recall (36.34) well above precision (24.88) for the leader.

Read carefully it is a bound rather than a verdict. Those tasks are consumer searches with a simulated user in the loop disclosing what they want, and nothing in them is concealed on purpose. Adversarial collection is harder on both counts, so an unattended profiler working a subject who resists should be expected to do worse than a third of what an expert would call complete.

Cross-spoke note

The agent machinery here (LLM orchestration, web-search tools, map-reduce over pages) is the subject of ../agentic-tooling-wiki; AI-OSINT is its application to intelligence-gathering. Filed here by subject, cross-linked there by mechanism.

Related

osint · llm-osint · kallisto-osinter · strix · autonomous-pentesting · social-engineering · synthesis