A campaign report arrives, and someone asks: Have we seen this here?
Our check of PoeLLM found two command-testing probes and a misleading malware-related filename. The probes came from visitors. The filename came from a directory listing generated by our own decoy. A search found both; reading the exchanges separated them.
That distinction is useful to any team turning a published threat report into an investigation.
Two probes worth following
On July 30, our decoys recorded two POST requests to /mcp-rest/test/connection. One supplied a stdio command configuration with a callback destination; the other supplied a simple command test. Both received 404 responses.
The two sessions contained 1,090 HTTP requests across a broader vulnerability sweep. These were attempts to test a command-capable interface, not ordinary visits to a home page. The next question is whether a real service accepted the input and did anything with it. Our captures show the submitted tests, not execution or a completed callback.

Why this campaign prompted a check
Lumen's Black Lotus Labs research describes PoeLLM deriving command-and-control addresses from a GitHub-hosted poem and identifies /mcp-rest/test/connection as a likely LiteLLM exploitation path. Lumen also traces its investigation to a compromised Ivanti Sentry system in early June that contacted the infrastructure and began scanning.
The report gave us concrete behaviors to search for. It did not make every matching path a PoeLLM incident. Our finding is an endpoint probe consistent with the exposure described; naming the campaign would require an independent link.
Run the check in four passes
- Start with behavior. Search relevant endpoints, methods, payload features, and outbound destinations. Check whether the affected software or interface exists in your environment. Preserve the original request rather than only the rule that matched it.
- Check direction. Did the visitor send the matched string, or did your system return it? Our
libgcryptmatch appeared in a decoy-generated directory listing. It was not a file delivered by the visitor. Keep analyst annotations and generated output separate from incoming evidence. - Expand the session. Read what happened before and after the hit. We also found
/private/bins.sh, another campaign-associated path, inside broad payload-path sweeps. Context distinguishes a generic dictionary match from a more identifying sequence. - Follow consequences. Look for related process creation, file writes, outbound traffic, or unauthorized workloads. A distinctive payload, matching hash, or corroborating command-and-control activity can strengthen attribution. An HTTP response alone cannot establish execution.
Give the responder evidence they can use
Every indicator hit should show its matched field, direction, surrounding exchange, and outcome. That turns a count into a decision: review exposure, investigate a probe, or contain corroborated execution.
Our research decoys retained the surrounding requests and generated replies, allowing us to keep the real probe and discard the misleading filename interpretation. That is the value of context: fewer false campaign claims, with the useful lead still intact.
Search for the indicator. Read the exchange. Follow what happened next.
Research notes
This retrospective was assembled October 7, 2026, from supplied live-log and archive material, including directly checked July 30 requests and their two sessions. No corroborated PoeLLM attribution was established in our reviewed captures. Source addresses, callback destinations, decoy identities, and executable request bodies are omitted. Campaign findings are attributed to Lumen and remain separate from our observations.
