Pentestas / help

Claude analysis

On Pro+ plans, Claude (Anthropic) participates in two distinct ways depending on scan mode:

  1. Post-processing analysis — narratives, false-positive filtering, attack chains (all AI-enabled modes).
  2. Workbench investigation — in + AI assist (hybrid), a bounded agent phase that does send carefully chosen requests using the same tools as the operator console. See AI workbench.

What Claude does (post-processing)

Given the raw findings, target context, and stack profile, Claude produces:

  • Plain-English narrative per finding — "Any authenticated user can read any other user's billing history by changing the account_id path parameter. This is a classic IDOR with no server-side authorization check."
  • Business impact paragraph — "Customer support ticket histories include credit card last-4, home addresses, and internal notes marking certain customers as VIP. Public exposure of this data violates PCI DSS 3.4 and likely your Terms of Service."
  • Remediation — stack-specific, code-level fix.
  • False-positive flag — if Claude concludes the finding is benign, it's either dropped or kept with ai_false_positive=true. See False-positive filter.
  • Attack chain synthesis — re-reads all findings together to find non-obvious combinations. See Attack chain synthesis.
  • Severity adjustment — raises or lowers severity by one level based on context.

What Claude does (hybrid workbench)

When Scan Mode is + AI assist and HIGH/CRITICAL leads exist (or an operator hint forces workbench):

  • Reads HTTP history and site map for the scan.
  • Sends custom requests via Forge / ranks Volley hits / mints OAST callbacks.
  • May pivot to Azure black-box recon (no spray) or AWS if keys are present.
  • Reports only with evidence — same proof bar as modules.

What Claude does NOT do

  • Invent findings. Every claim must reference real evidence (module or workbench traffic). The output is not free-form hallucination.
  • Spray blindly. Post-processing does not open new traffic. The workbench phase is targeted investigation, not a second full-module run.
  • Train on your data. Per our contract with Anthropic, customer data is not used for model training. Inputs are sent with the no-training flag.

Bring-your-own-key (Pro+)

You can use your own Anthropic API key:

Settings → AI → Anthropic API key.

The key is encrypted with your tenant's Fernet key and only decrypted for the duration of the AI pass. Usage counts against your own Anthropic billing, not ours. Useful when:

  • You have a negotiated enterprise rate with Anthropic.
  • You require all AI traffic to use your own account for compliance.
  • You want to hit higher rate limits than the default pool.

Without a BYOK, Pentestas uses a shared pool. Rate limits apply (typically 200 findings per scan can be AI-analysed; the rest fall back to rule-based narratives).

Which findings get analysed

  • All CRITICAL + HIGH findings always.
  • MEDIUM findings if the total finding count is under the per-tenant cap.
  • LOW/INFO findings — skipped by default; surface in the report only with rule-based narratives.

Turning it off

Settings → AI → Enable AI analysis → uncheck. All findings still flow, just without the narrative + remediation + chain synthesis. Rule-based output (template narratives, default remediation) is still produced.

See also