We also give credit to the rest of the authors of our report, including co-leads Jonathan Elsworth Eicher and Rowan Howard-Jones. This release was mostly prepared by Concurrent, using a combination of Concurrent's own and Ethan's archived data.
We’re releasing our records related to the Tencent fleet we reported on last month, covering over 3,000 URLQuery reports, 775 unique httpbin payloads, 377 ntfy.sh messages, 627 mailbox messages, and the 78 messages sent between us and the fleet. We are doing this to help improve the world’s understanding of AI fleets and swarms. Such understanding would allow defenders to better identify and quickly respond to future fleets and swarms.
Our dataset contains 10 nice files, chief among them being:
reports.jsonl: a list of 3,000 URLQuery reports related to the Tencent fleet we’ve identified, spanning the 8 days between September 30 and October 8. Each report includes its scan time, each request it made, and also a reference to a httpbin payload in payloads.jsonl, if there is any; along with many other nice metadata fields.
ntfy-messages.jsonl: partial archives of 2 nfty topics related to the fleet. Our archived nfty data includes a full run of an agent successfully obtaining an Amap place on its own, along with the full payload (of answers) it received.
webhook-requests.jsonl: a partial archive of 23 (out of the 32) mailboxes containing 628 messages from the fleet.
mailbox-conversations.jsonl: 78 messages (47 from agents, and 31 from us), comprising the world’s first communication with an AI fleet. See the report for more details.
Some information - specifically the API keys and security codes used by the fleet - are redacted. While we believe almost all of the API keys are invalid or have been otherwise deactivated, we don’t want to leak any potentially valid API keys. Additionally, we don’t think the unredacted API keys or security codes are required for our actual goal: helping people build classifiers for future swarm traffic.
We hope that this data can help others achieve this goal. Empirically, we have found that our coding agents are sometimes ineffective at detecting it without many few-shot examples, and we would be interested in work that tries to build cheap classifiers that we can set on many websites.
We’re releasing our records related to the Tencent fleet we reported on last month, covering over 3,000 URLQuery reports, 775 unique httpbin payloads, 377 ntfy.sh messages, 627 mailbox messages, and the 78 messages sent between us and the fleet. We are doing this to help improve the world’s understanding of AI fleets and swarms. Such understanding would allow defenders to better identify and quickly respond to future fleets and swarms.
Our dataset contains 10 nice files, chief among them being:
though a few other useful files are in the rest of our repository on Github. We may also add more archived data in the future.
Some information - specifically the API keys and security codes used by the fleet - are redacted. While we believe almost all of the API keys are invalid or have been otherwise deactivated, we don’t want to leak any potentially valid API keys. Additionally, we don’t think the unredacted API keys or security codes are required for our actual goal: helping people build classifiers for future swarm traffic.
We hope that this data can help others achieve this goal. Empirically, we have found that our coding agents are sometimes ineffective at detecting it without many few-shot examples, and we would be interested in work that tries to build cheap classifiers that we can set on many websites.