The investigators added offscreen text to a report they controlled, inviting apparent automated readers to contact a mailbox. They subsequently recorded 33 messages across 13 threads and say 12 investigator responses were read. The important boundary is not the mailbox itself: content returned by a lower-trust page appears to have influenced what a reader…
Inspect the ClaimsCase · DiggingBeagle record
Hidden text in research report redirects observed agent communications
The investigators planted an offscreen invitation on their own page; some apparent agents contacted a mailbox and exchanged messages, demonstrating a lower-trust document influencing subsequent tool behavior. This is a distinct observer-origin intervention, not the fleet's original map query method.
Scope
Investigator-designed intervention against public-document ingestion
Assessment profile
Assessment method- Setting
- controlled live
- Exploitation
- observed live
- AI role
- AGAINST AI
Basis and provenance
Assessment is limited to the selected canonical Claims and their Source locators in knowledge revision 152. Primary public URL retrieval was unavailable during this drafting pass, and the pinned index reports incomplete linked-record coverage without identifying missing IDs. No independent production impact, generalized exploit prevalence, or attributable victim count is inferred. Case setting is controlled_live; evidence status records only the documented source-reported result, not external replication.
Assessed Oct 15, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-2026-AGENT-FLEET-HIDDEN-TEXT-01 · Swarmchasers: full technical analysis of Tencent-hosted agent fleet
At a glance
Attribution of every thread to a specific agent, controlled comparison, and any privileged side effects beyond voluntary messages are not established.
Limits and uncertaintyFull account
The investigators added offscreen text to a report they controlled, inviting apparent automated readers to contact a mailbox. They subsequently recorded 33 messages across 13 threads and say 12 investigator responses were read. The important boundary is not the mailbox itself: content returned by a lower-trust page appears to have influenced what a reader operating as an agent did next.
This intervention was created by the investigators after observing the fleet; it must not be confused with the fleet's original public-map research and scanner use. The communications support a narrow finding of behavioral redirection, not privileged command execution, secret extraction or reliable attribution of every correspondent to the same orchestrator. The relevant comparison is the separate remote-browser fleet Case, where the primary issue is use of a third-party scanner to run submitted browser content.
Claims & evidence
CLM-2026-AGENT-FLEET-HIDDEN-TEXT-01The investigators report agent contact after adding an offscreen mailbox note to their report and recorded 33 agent messages across 13 threads, with 12 investigator responses read.supported
Basis: reported finding
- supportsSwarmchasers: full technical analysis of Tencent-hosted agent fleetprimary disclosure
Communication with fleet, hidden-text note added 7 October
Unknowns and contradictions
- Attribution of every thread to a specific agent, controlled comparison, and any privileged side effects beyond voluntary messages are not established.
Sources and citation
Material revision history
- Oct 15, 2026 · Canonical change recorded · new in release · revision 153