Fake AI bots scan organisations for exposed credentials

A Cequence Security Cequence CQ Prime Research report details a threat actor that operated behind more than 30 AI bot identities to probe organisations for exposed credentials. The malicious scanner impersonated 32 bots—including crawlers and assistants attributed to OpenAI, Anthropic, Meta and Perplexity—and sought files containing passwords, API keys and cloud credentials.

The activity has been ongoing since late September and remained active into late October; researchers observed that on 21 October nearly half (48%) of Cequence-protected organisations were targeted by the bot traffic in a single day. The scanning tactic involved rapidly sending 50 to 190 requests in about eight seconds, then moving on to the next target, with the vast majority of requests either blocked or failing.

CQ Prime Research highlighted several telltale indicators that the traffic was malicious rather than legitimate bot activity. The scanner rotated through about 32 bot identities in a way that normal bots do not, and one machine impersonated multiple Claude agents by citing Anthropic’s crawler contact address for all Claude identities. It also requested configuration files, cloud credential files and private SSH keys from roughly 90 paths while withholding any contact information.

While the activity could enable access to accounts if successful, Collin Hogue-Spears of Black Duck emphasised that the findings describe a test, not a breach. Financial institutions feature prominently in concern, with analysts noting that adversaries may already be logging in using customers’ credentials via ordinary browsers.

The report suggests security teams should treat AI agents, API keys and service accounts with rigorous validation, verify agents through signed requests or direct integrations, monitor for identity-abnormal behaviour, and move away from static credentials toward centralised, zero-knowledge secret vaults.