AI news · October 16, 2026
Researchers report more than 2,600 AI-agent attempts against Amap data
A report counted more than 2,600 AI-agent attempts to extract Amap data in 11 days, so agent builders need destination allowlists, key rotation and logs before a tool can browse freely.
- reported attempts
- 2,600+
- reported window
- 11 days
- agent inboxes observed
- 17
Swarmchase published a report on October 14 describing more than 2,600 attempts to extract Amap data between September 28 and October 8. It says the attempts included bypasses, exposed keys and preproduction systems, with 17 agent inboxes hosted on Tencent Cloud. Most attempts failed, but the report says some data was retrieved. It found no evidence that one coordinated actor ran the fleet and said the model or operator remains uncertain; Tencent had not officially responded in the report.
The report is not proof of a single attack or of a successful compromise, but it shows how a large number of agent runs can turn an ordinary data task into repeated requests against a live service. The useful response is operational: limit destinations, isolate keys, log requests, rate-limit retries and make unusual access stop. A product should record enough evidence to tell a bad model decision from an abused credential.
What you can do with it
Read the report, then run a closed test with synthetic data and a deny-by-default destination list. Rotate the test key, log every request and force a stop after repeated failures. Check that a human can identify the model, tool, destination and credential used for each action.
Our take
The incident is a warning about scale and visibility, not a reason to ban agents. The uncertain operator and lack of official confirmation limit the claims, while the exposed-key pattern is concrete. Make request logs and destination controls part of the first build, not a later security task.
Links InvestigationCoverage
Source: Swarmchase — Made With Models writes the brief; the reporting is theirs.