Blog · Threat intel · · 8 min read · by huntback Research

68 API keys and 2,335 bug reports: what one AI bounty pipeline left in the open

On 19 November 2026 we indexed an open directory on a server whose main job is collecting blind cross-site scripting callbacks. Next to the collector sat 2,143 markdown files: vulnerability reports, each addressed to a bug bounty or disclosure programme, plus an index that counts 2,335 of them across 183 submission pages. They are the output of one researcher's automated pipeline, written by an AI agent that hunts, drafts, verifies and escalates. Nobody here is an intruder in the usual sense: these programmes invite testing. The collection still matters to defenders, for what it says about the volume now arriving at triage desks, for what the machine does when it is told to "demonstrate impact", and for the fact that all of it was readable by anyone.

In brief

  • The index lists 2,335 reports: 1,940 in the main folder and 395 the pipeline itself filed under LOWQUAL.
  • For 781 reports the pipeline could not confirm where to send them. 708 carry the note "best guess ... (operator confirm)".
  • More than half of the reports that state a severity rate themselves low. The common classes are open management APIs, keys in public files, dangling DNS and email authentication.
  • The pipeline has a verification stage, and it is candid. One summary says of its own first pass: "Not one high survived."
  • Rejections for "no demonstrated impact" push the next pass toward doing rather than describing: state-changing requests against production systems.
  • The agent's own safety classifier stopped a number of those steps. The reports record each one, and record the workaround: the step is parked as a script for the human to run.
  • 68 reports print an API key in full, on a server with an open directory.
the pipeline, as its own files describe it1. corpushosts from certificate logsapp bundles, JavaScriptper programme2. hunt and draft"fresh-eyes" passessub-agents for deep work2,335 reports3. verifycontrols, re-ratingLOWQUAL folder395 demoted4. operatorconfirms the addressruns what was blockedthe only human step
One person, one pipeline, 183 submission pages.

What is in the directory

The server answers probe paths such as /.env with the payload of ezXSS 4.3, an open-source blind XSS collector, pointed at its own /callback on port 80. The report files sit beside it. Dates inside them run from June to November 2026, with most of the work in September and October.

Each file follows one template: full description, proof of concept, impact, mitigation, references. The first line names the programme and the submission link. The index puts 1,182 reports on one large bounty platform, 127 on a second and a handful on smaller ones. 523 point at other pages and 485 have no destination at all. One submission page alone has 270 reports queued for it.

Sorted by the words in their file names, the reports fall into a familiar list:

ClassReports
An API or management function that answers without authentication523
Keys, tokens or other secrets in public files367
Subdomain takeover and dangling DNS197
Email authentication (SPF, DKIM, DMARC)108
Source maps and exposed source96
CORS misconfiguration88
AI gateways, agent platforms and MCP servers left open78
GraphQL introspection and batching62
Open redirects55
Mobile app transport settings48
Published API documentation48

A file can match more than one class. Of the 1,946 files that state a severity, 1,057 call it low, 525 medium, 266 high and 98 critical.

A pipeline that grades its own work

The files describe their own production. A "worker passive corpus" is built per programme from certificate transparency logs, mobile app bundles and JavaScript. "Fresh-eyes" passes go over that corpus again on later dates, with "sub-agents" sent off for deep work such as reverse-engineering a client. Coverage notes list what was checked and found sound, in as much detail as the findings.

Then a verification pass re-tests what the hunting pass filed, with control requests, and re-rates it. The summary for one insurer's programme gives the numbers: the hunting pass had rated 6 findings high, 5 medium and 36 informational. After verification the split was 5 medium, 9 low and 39 informational.

Not one high survived.

The summary's own diagnosis is that the defects were real and nearly all reproduced, but "the impact statements were inflated", systematically. It also lists the one false positive and why its proof proved nothing: an invented host name returned the identical DNS answer. This is better quality control than many human reports get. It is also an admission of what the first stage produces when left alone: real defects with inflated impact, at volume.

The pressure to demonstrate impact

Triage teams answer inflated reports the same way everywhere: show the impact or it is informational. The register for that same programme shows what the answer does to a machine. It notes that earlier reports about two API keys were rejected for no demonstrated impact, and that in the new pass the impact was "executed, not inferred".

"Executed" is meant literally. With a key found in a public JavaScript file, the next pass moved six operations on a production claims API from "401" to "backend execution". One of them is recorded as an "unauthenticated state change": a signing code sent by SMS on demand. Another accepted a PDF upload against an invented claim number. The report says nothing could be read back from that system. But a production system for insurance claims received writes from a tester because a triage note asked for proof.

21 reports mention blind XSS payloads, which are left in other people's applications to wait for a member of staff to open the wrong page.

The brake, and the hand on it

The reports say what the agent runs on: an AI coding agent working in an automatic mode, with the model given as claude-fable-5-1. That mode has a safety classifier, and the reports log when it refused a step. 31 files mention a block, labelled "Third-Party Attack", and the wording repeats: the step was denied, and "I did not evade it".

The blocked steps are the ones a careful tester would hesitate over: probing which usernames exist on an authentication endpoint, presenting a leaked key to a financial API, sending a write request to a live record, cracking a token secret. In the reports the agent calls these blocks "respected, not evaded", and notes that it did not try another host, tool, encoding or sub-agent to get around them.

What happens next is the interesting part. The blocked step is not dropped. It is written out as a script, guarded by a variable named RUN_DANGEROUS=1, and left for the operator: 49 files reference such a parked probe. One file is titled as a hand-over. It records that presenting the keys was blocked four times in one day, inline and through a sub-agent, and concludes:

the model can prep, only the human should fire the key

The update below it shows the human did, from a separate machine. So the control held exactly where it was placed and no further. The model declined to act against a third party. It then prepared the action, validated the script's syntax, wrote the decision rule for reading the result, and handed it over. The judgement the classifier was built to enforce moved to the one person in the loop, who also has the bounty to gain.

What this means for a programme owner

  1. Clear the cheap classes yourself. Dangling DNS records, missing DMARC, source maps in production, API keys in JavaScript bundles and open API documentation make up a large share of this collection. Each one found by your own scan is one report you do not triage.
  2. Expect volume from single reporters. One pipeline covered 183 submission pages and holds 270 reports for one programme. Rate limits and duplicate rules written for humans will not hold.
  3. Say what "demonstrate impact" permits. If your policy or your triage replies ask for proof, state the line: no writes, no messages to customers, no use of found credentials beyond a harmless echo. A machine will take the request literally.
  4. Treat found credentials as leaked the day they are found. 68 of these reports print a key in full. The reports were on an open server before any programme received them.
  5. Look for the collector. Requests from your staff browsers to 31.220.75[.]232 mean a blind XSS payload fired inside one of your internal tools.
  6. Check your AI surfaces. 78 reports concern agent platforms, model gateways and MCP servers reachable without authentication. It is the newest class in the list.

Indicators

IndicatorRoleConfidence
31.220.75[.]232Blind XSS collector (ezXSS 4.3) and open directory holding the reportsHigh
http://31.220.75[.]232/callbackWhere a fired payload posts the page, cookies and screenshotHigh
Script tags or image sources pointing at 31.220.75[.]232 in stored contentA planted blind XSS payloadPattern