A scanning campaign is working its way across the web disguised as AI crawlers, and what it wants is specific: the configuration files that AI coding tools leave on disk. Known Agents, which measures bot traffic across more than 5,000 websites, lists it as an active threat and says the attacker is targeting credential and configuration paths used by AI coding tools.
The paths being requested read like an inventory of a developer machine. Among them are /.claude/settings.json, /.codex/config.toml, /.aider.conf.yml, /.aws/credentials and half a dozen variations on .env, including .env.production and .env.backup.
What the scan is looking for
These are not exotic targets. They are the files AI coding assistants write when you configure them, and they routinely hold API keys, model endpoints and tokens for whatever the tool was given access to. A file such as /.config/anthropic/credentials/default.json is only dangerous if it is reachable from the public web, which is exactly what the scan is testing.
The campaign also probes /service-account.json, /firebase-service-account.json, /terraform.tfstate and /rclone.conf. The common thread is that each one, if served, hands over working access to something, rather than merely confirming that a system exists.
Why the disguise works
Bots account for roughly 35% of the traffic Known Agents measures, and 29% of that bot traffic is AI related, a share that grew by 11% over the previous 90 days. Traffic claiming to be a well known crawler is ordinary enough that it does not stand out, and plenty of sites wave it through on the user agent string alone.
That is the weakness being used. Known Agents counts a visit as spoofed when it claims a recognised agent identity but fails that agent’s own authentication, such as a verified IP range or Web Bot Auth. Its ranking puts Googlebot as the most impersonated identity at 0.5%, with ClaudeBot, GPTBot and PerplexityBot each at 0.1%.
Robots.txt offers nothing here. Known Agents measures 98.5% compliance with it, but that figure describes bots that identify themselves honestly. A scanner wearing someone else’s name has already decided not to play along.

Why Australian businesses should care
Any Australian business that has put AI coding tools in front of its developers now has this class of file somewhere. The risk is not the tool. It is a deployment that copies a project directory onto a server without stripping what was only ever meant to sit on a laptop.
If one of those files is served and the keys inside reach personal information, it stops being a technical problem. The OAIC requires eligible data breaches to be reported, and an exposed credential granting access to customer records is the sort of thing that qualifies. The organisation carries that obligation, not the tool vendor.
It also runs straight into the machine identities question we covered earlier today. The keys sitting in these files are those identities, written to disk. This campaign is someone checking at scale who left them somewhere reachable.
What to check this week
Request the paths against your own public hosts and see what comes back. Anything other than a 404 deserves attention. The exercise takes minutes and needs no tooling beyond curl and the list itself, which is published openly.
After that the work is unglamorous. Keep dotfiles out of the web root, verify crawlers by IP range or Web Bot Auth rather than trusting the user agent string, and rotate anything that has ever been committed to a repository or shipped to a server. That last point matters because deleting a key from a file does not delete it from the history or the backups.
The web is not just people now
There is a legitimate side to this traffic. Real AI search crawlers visiting to index content are a growing source of referrals, and blocking everything that looks like a crawler costs visibility. Telling the two apart has become an operational skill rather than a niche one.
For teams running local AI agents or coding assistants across a fleet of machines, the question is narrower than it first appears. Not whether the tools can be trusted, but whether their configuration has ever left the machine it was written on.
