Skip to main content

Autonomous Pentest

Path: /scans · Sidebar: DAST → Autonomous Pentest

apPosture's Autonomous Pentest behaves like a human penetration tester: it sends a probe, reads the response, and decides the next payload - instead of replaying a fixed signature list. You watch it work in the live Agent Cockpit, and every confirmed finding leads with an Attack Trace - the proven, step-by-step kill chain - rather than a bare CVSS number. Every verdict is backed by deterministic proof.

How it works​

  • Adaptive, agentic loop - probe → observe → mutate → escalate. A blocked payload is mutated; a dead parameter is swapped; a quote-error is escalated to a timing or out-of-band proof.
  • Cross-finding chaining - the engine feeds one primitive forward instead of testing endpoints in isolation. A JWT or API key leaked to an anonymous reader is replayed against privileged endpoints (credential propagation); a reference in a response body is followed to an endpoint the crawl never reached (object-graph IDOR); a confirmed SSRF is walked from the AWS metadata service through IAM role enumeration to usable AWS credentials. This is what "autonomous" means concretely - it builds the multi-step path a single-endpoint probe misses.
  • Deterministic adjudication - the LLM drives the search, but a deterministic adjudicator (marker / timing / canary / differential) gives the verdict - so it's "here's the proof", not "the model said so". False positives drop sharply.

Proof, not guesses​

A finding is raised only when an exploit or leak is confirmed by an oracle, not a heuristic. The confirmation tiers include:

  • Injection - arithmetic-eval and unique markers; boolean-blind, time-blind and a fused boolean+time oracle that ties a boolean predicate to the timing channel (a false predicate is a built-in negative control against uniformly-slow endpoints); out-of-band (OAST) callbacks for fully-blind sinks.
  • Second-order / stored - a payload stored at endpoint A that fires at a different consumer endpoint B, confirmed out-of-band, including a delayed re-check that captures a callback landing minutes to days after the scan.
  • Access control - strict differential (denied without the artifact, granted a substantial and distinct body with it) for credential propagation and object-graph IDOR; 2-user checks on header-authenticated APIs.
  • Business logic - stateful race conditions and financial-tamper workflows proven by reading application state back.
  • Behavioral differential discovery - ruleset-agnostic injection reach detected from response behaviour, not a signature match.
  • Global-reflector pre-flight - a site-wide false-positive killer: a target that echoes any input on any path is detected and its unconfirmed reflection findings are demoted, so mock/tarpit servers do not flood the results.

Safety (dual-use restraint)​

The engine is for authorised testing and is non-destructive by design:

  • benign markers only - no real exploit, no gadget execution, no destructive verbs (DELETE/PUT/transfer); the live view carries a visible SAFE trust signal while these checks run;
  • host-pinned, read-only, with a request budget;
  • exfiltrated secret material (e.g. an AWS key) is masked in the proof;
  • live/production targets require confirmed authorisation; no DoS / bursting.

Running it​

Start a scan and pick a profile (quick / standard / deep); the autonomous pentest phase runs after crawl + active scan and feeds proven findings into the posture. Watch it live in the Agent Cockpit, then read the proof-of-exploit and Attack Trace on each finding.

The Autonomous Pentest report​

Each run can be exported as a vendor-grade Autonomous Pentest report - a per-scan deliverable modeled on a professional penetration-test document:

  • Project objectives, Scope & timeframe, and a Security Grade (A-F).
  • Summary of business risks and high-level recommendations.
  • Compliance status by framework (a framework passes at ≥80% control coverage).
  • Appendix A - status against the OWASP Top 10 (2021): a per-category pass/fail table.
  • Methodology mapped to PTES / OWASP WSTG / ASVS, and the tools used.

It is scoped to the one scan it documents. See Report Types for the full structure and export formats.

Last updated on