Back to blog

Picus Autonomous Pentesting: What Independent ACOST Validation Must Prove

4 min read
Picus Autonomous Pentesting: What Independent ACOST Validation Must Prove

Sources checked: . FailSafe Research.

Agentic pentesting is now a product category, not a differentiator. Picus launched Autonomous Penetration Testing on 12 November 2026. Cisco expanded automated AI red teaming in its security platform. CrowdStrike entered the previous week. The buying question has moved from whether a vendor has agents that chain exploit steps, to what evidence a vendor can produce that a specific path was reproduced and then closed.

What Picus announced

Picus describes AI-agent swarms that reconnoitre an environment, attempt real exploits, adapt attack paths toward critical assets and validate defenses. Customers predefine scope and permitted techniques. Hard-deny rules and an external kill switch constrain execution. The company says existing functionality is available to its customers, while full-agentic operation remains in early access. Those are vendor descriptions. No independently measured comparison of this release is published here.

The structural advantage is real. Picus brings an installed breach-and-attack-simulation and exposure-validation customer base, several years of automated pentest deployments, control-validation integrations and an enterprise partner position. That makes it a closer competitive comparison than a generic AI red team or a runtime-security product.

The comparison that matters is the evidence boundary

Two platforms can both say "chains real exploits" and still produce different things. Ask what the customer receives after a successful chain. A screenshot of a terminal is not a record that connects the entry condition, the authority used, the observed effect and the control that should have intervened.

Questions that separate agentic pentesting offerings
DimensionQuestion for the vendorWhat a usable answer includes
Target scopeDoes one engagement cover applications, APIs, cloud, identity, AI systems and transaction workflows, or separate products?A single attack graph with the authority path between assets.
AuthenticationCan the test operate with valid credentials and delegated roles?Named test identities and the permission boundary used.
ImpactIs the consequence reproduced, or inferred from a request?A bounded synthetic effect and the observed result.
Control attributionWhich control interrupted the path, and did it log the decision?Correlated application, identity and control records.
Evidence qualityCan an engineer who was not present repeat the case?Version, configuration digest, fixture and attempt budget.
Safety controlsWhere does the stop rule execute, and who owns it?A control outside the test agent, plus a tested stop case.
RetestIs closure verified on the deployed build?Repeat of the original case with the legitimate control.
IndependenceIs the assessor the same party that sells the control under test?A stated relationship between assessor and control vendor.

Platform bundling is the second pressure

Cisco announced that AI Defense will use Anthropic's Claude Mythos for adaptive red teaming of AI applications and agents, combined with Splunk agent observability and Cisco infrastructure intelligence. Cisco has not published comparative accuracy or coverage results for that configuration. The significance is distribution. Another large incumbent can attach agent testing to networking, identity and observability relationships instead of selling a separate product.

Vanta's agreement to acquire Pillar Security points at a third route: adversarial testing sold with compliance evidence. Each route prices testing differently. None of them removes the need to demonstrate that a particular attack path reached a particular consequence in the customer's environment.

How to run a matched evaluation

Buyers can compare offerings without claiming a winner from marketing pages. Authorize one representative environment. Give each provider the same scope, test identities, permitted actions, attempt budget, stop rule and evidence format. Grade the outputs against the same rubric, and keep blocked and inconclusive cases in the record. A provider that reports only successes is not giving you a comparison. A provider that reports only failures may be running a narrower scope.

Include one case that should be blocked by a strong control. If neither provider can show the control decision, the test measured reachability, not defense. Include one legitimate operation that must keep working, so a fix that breaks the business is visible.

Where FailSafe fits

FailSafe SWARM runs authorized attack-path testing across applications, APIs, cloud, identity and AI systems, with reproduced findings and remediation retesting. The agentic pentest buyer guide lists twelve evaluation questions, and the reproducibility protocol describes how to judge a finding. A sovereign deployment keeps model processing and assessment evidence inside an agreed boundary.

We are not publishing a benchmark against this release. A dated, separately defined CVE-Bench comparison is at the benchmark report. Use a matched evaluation on your own environment for a purchasing decision.

What this analysis does not establish

Picus performance, production safety and full-agentic behavior remain vendor-reported and partly early access. This article evaluates the claims buyers can test, not the product's measured effectiveness. Product scope and availability change, so confirm terms directly with each provider.

Questions & answers

Questions about this topic

Direct answers to common questions about this page.

Picus says existing functionality is available to its customers, with full-agentic operation in early access. Confirm the current scope and terms with the vendor.

Yes, by authorizing one environment with the same scope, identities, permitted actions, attempt budget and evidence format, then grading each output against the same rubric.

No. This article compares testable evidence requirements. No matched evaluation of this release is published here.

An assessor that also sells the control under test has a different relationship to the evidence. Independent validation is about who observes and reports the result.

Put this guidance to work

Bring one representative environment, the identities and permitted actions, and the evidence your team needs to compare providers.

Startup offer: $999 USD for 250 credits and one authorized black-box run. Sign in to SWARM before Stripe checkout. Custom scope and sovereign deployments are separate.

Plan a matched agentic pentest evaluation