Adverse-agent intelligence
Tracking down harmful AI agents and agent swarms.
More than 21 incidents involving rogue agents from different labs have become public. In one 2026 campaign, an attacker directing AI agents hit 440 exposed servers in 48 countries and compromised 11 organizations. Only a third of organizations give their agents least-privilege access.
Sources: Pillar Security, Plain English, GreyNoise, Cequence and EMA, DataDome
- SensorUnknown agentUnclear
- Partner siteSearch crawlerBenign
- Public sourceUnknown agentUnclear
- Partner siteAI assistant fetcherBenign
- SensorData fetcherCareless
- Partner siteUnknown agentUnclear
- Our labUnknown agentRogue 0.71
One agent, followed across sources
- Verdict
- Rogue · 0.71
- Fingerprint
- fp:7c1e…a90b
- Infrastructure
- AS64512 · cloud host
- Seen
- 02:14 → 02:40 · 4 sources
- Swarm
- None, acts alone
- Techniques
- T1595 · T1213
The operator gets a private notice first, with the evidence.
An illustration of how AgentRunnr will work, not real data.
Agents pose a different challenge.
AI agents raise questions traditional threat intelligence wasn’t designed to answer.
What was it after?
Agents pursue tasks. They hunt for specific data and adapt when blocked, which can look nothing like a scanner. We monitor for this kind of behavior.
How was it run?
A rogue agent, a carelessly deployed agent or one directed by an attacker can look identical. We can help identify which, including benign agents.
Who's was it?
Behind each agent is a model, a framework and an operator. When the evidence allows, we work out what is behind each agent, and notify the people who can stop it.
We build a case across many sources, following the same agent or swarm from target to target.
- ObserveCollectEvery source feeds one store of raw, timestamped evidence that nobody can edit.
- OrientFilter and analyzeAI-assisted triage that cites evidence.
- DecideConfirmA human signs off on every finding and notice.
- ActFindingVerdict, attribution, swarm, sightings and evidence, sent to the people who can stop it.
Each finding makes the next one faster.
Evidence driven.
Confirmed findings carry a confidence level and evidence that can be checked
- Identification
- Benign, careless, rogue or directed by an attacker, with a confidence score, or unclear until the evidence settles it.
- Rogue · 0.71
- Attribution
- Whose it may be, where the evidence allows: model, lab or infrastructure, never individual people.
- AS64512 · cloud host
- Swarm
- Whether it acts alone or is part of a swarm or wider campaign.
- None, acts alone
- Sightings
- Where else it has been seen, across targets and sources.
- 4 sources · 02:14 → 02:40
- Evidence
- What it did: timestamps, sources, tools, messages and the surfaces it hit, kept raw and checkable.
- T1595 · T1213
Finding confirmed
- Day 0
Private notice
The suspected operator hears from us first, with the evidence.
- Fixed period
Response window
A fixed window for the operator to investigate, respond or act.
- After the window
Affected parties told
Victims are always informed, whoever the operator is.
- Where appropriate
Public findings
We publish and share findings through open standards.
That’s the usual order. If agents are actively causing harm, we don’t wait for the window. If a person appears to be directing agents to commit crimes, we start with the people affected.
We never engage agents. We alert the people who can stop them and help targets protect themselves.
When people appear to be directing agents to commit crimes, we alert victims and hosting providers first and refer to law enforcement where appropriate. We only identify the tools and infrastructure involved.
Who we're working with first.
During the alpha, we’re hoping to work closely with a small number of teams in each of these groups.
Data and API teams
Find out which agents came for your data, with an Agent Exposure Check of your past logs.
- Which agents reached your data, and which got around your controls
- A verdict for each, where your logs show it: benign, rogue, careless or directed
- A blocklist, then a daily digest for alpha members
Security platforms
Get the context you can’t see from your own traffic.
- Indicators, fingerprints and campaign reports to build into your products
- Send us agent sessions you’ve detected and get a verdict back
- Whose each agent was, where the evidence allows
Free for every site
Share your access logs with us and see which AI agents visited your site.
- A weekly summary of known and unknown agents
- A heads-up if we see harmful activity
- Free, and it stays free
Agent Exposure Check
api.example.com · logs from July 1 to September 30
- Agent sessions
- 48,906
- Agents found
- 27
- Came for your data
- 11
- Got around your controls
- 4
| Agent | Verdict | Sessions | How it got in |
|---|---|---|---|
| Unknown agent C | Rogue | 6,412 | Rotated keys after rate limits |
| Unknown agent D | Directed | 1,980 | Used an old, unauthenticated endpoint |
| Data fetcher E | Careless | 1,207 | Ignored robots.txt and paging limits |
| Unknown agent F | Unclear | 388 | Spread across many cloud hosts |
Plus a blocklist of 31 hosts and fingerprints, and the evidence behind each finding.
Weekly agent summary
example.com · October 19 to 25
- Agent sessions
- 1,284
- Known agents
- 9
- Unknown agents
- 3
- Heads-ups
- 1
| Agent | Verdict | Sessions | Most visited |
|---|---|---|---|
| Search crawler | Benign | 812 | /docs/ |
| AI assistant fetcher | Benign | 301 | /pricing/ |
| Unknown agent A | Unclear | 97 | /api/datasets/ |
| Unknown agent B | Careless | 74 | /login/ |
And 8 more agents in the full summary.
Heads-up: Unknown agent B
Tried your login page 214 times in six minutes from three cloud hosts. Verdict: likely careless, with no sign of a person directing it. Consider rate-limiting /login/.
Join the alpha.
For data and API teams, security platforms, AI labs, researchers, insurers and government teams. Agent Exposure Checks and weekly agent summaries will come with the alpha launch.
You’re on the list
Thanks. We read every signup ourselves and will reply by email to . The alpha opens in small groups.