loading
Loading content
loading
Scanners
Black Hat Arsenal auditing for web-facing AI agents through browser automation, with no provider token.
overview
webagentaudit tests the AI agents that companies embed in their own web pages: the support chat, the shopping assistant, the onboarding bot. Those agents have no API you can point a scanner at. They exist only behind a widget, so the audit drives a real browser, finds the chat surface, and interacts with it the way a visitor would.
Mode selects how far it goes. detect looks for the chat surface and reports where it is. assess runs the algorithmic probes against a surface it has found. fixture replays a demo for shape checks. Because the probes are algorithmic rather than model driven, no provider API token is required, which keeps the node runnable on a schedule without a per-run cost.
Targets arrive as a single --url or a file on the url-file port, and the Web AI Agent Audit workflow probes hosts with httpx first so only live pages reach the browser stage. Output is findings.jsonl plus screenshots and reports, so an assessment leaves evidence a retest can be diffed against.
use cases
Run detect mode across a URL list to inventory the chat surfaces on your estate, including the ones a team shipped without telling anyone.
Switch to assess mode so the browser exercises the widget with algorithmic probes and records what the agent does in response.
Screenshots and reports land beside findings.jsonl, so the next scheduled run can be compared against what the agent did last time.
Chain httpx before the audit, as the Library workflow does, so browser automation runs against pages that resolve instead of the whole list.
reference
| Name | Type | Flag | Description |
|---|---|---|---|
| url | STRING | --url | Single chat-page URL (optional if the urls file is set) |
| urls | FILE | --url-file | URL list file (one per line). Prefer seed to prep wiring, distribute per line. |
| mode | STRING | --mode | detect (find chat widget) | assess (run probes) | fixture (demo only) |
Showing key inputs. webagentaudit exposes 3 inputs in total.
example
# find the chat surface on one pagewebagentaudit --mode detect --url https://www.example.com/support # probe a list of pages you are authorized to testwebagentaudit --mode assess --url-file urls.txt{ "tool": "webagentaudit", "url": "https://www.example.com/support", "mode": "assess", "surface": "chat widget detected", "probe": "system-prompt disclosure", "result": "agent returned configuration text", "severity": "medium"}guidance
Reach for webagentaudit when the agent is reachable only through a browser widget on a page you are authorized to test. Use mcparasite when the agent speaks MCP and the question is cross-channel propagation, and use httpx when you only need to know which pages are live.
Tests agents across MCP propagation channels. Same class of risk, reached through a protocol client rather than a browser.
Probes hosts and reports what responds. It finds the pages; it cannot interact with a chat widget rendered by JavaScript.
Templates against the underlying web app. Runs beside an agent audit and covers the ordinary web surface, not the agent behaviour.
workflows
faq
related
BHUSA Arsenal TLS/SSH crypto inventory with post-quantum readiness signals and CBOM-shaped output.
Reflection triage for parameter URLs during recon.
Black Hat Arsenal hunting for phantom IAM users behind Bedrock keys, with offline key decode and org scan.
CMS detection and version fingerprinting.
Signature-driven web application scanner.
OWASP Joomla vulnerability scanner for CMS flaws and misconfigurations.
A list of live page URLs feeds webagentaudit, which detects chat surfaces, probes them, and writes findings with screenshots.
Facts on this page come from the live Trickest tool library.