Skip to main content

Detectify's Agentic Tooling lets AI agents verify flaws in your environment

Ilustrační obrázek
Detectify says roughly 80% of what AI-driven static analysis flags is noise. On 16 November the Swedish vendor made Agentic Tooling available on request: a customer's own AI agents test each suspected flaw against a running application, and an append-only Security Ledger records every request, payload and result. Detectify is not selling better bug detection. It is selling evidence that survives an audit.

That difference matters. Most security announcements in 2026 add another scanner to the pile. This one adds a verification step and a receipt for it.

What the tool does

DevSecOps teams already sit on a queue of suspicions: SAST findings, bug bounty reports, notes from an internal red team. Agentic Tooling lets a company's own AI agents take each suspicion and test it against the running application, rather than against a snapshot of source code.

The agents run in the customer's private environment and connect through the Model Context Protocol or a CLI. Detectify states that it neither stores nor accesses customer codebases. Requests leave the environment; code and payloads stay put.

That architecture is the whole product. An agent probing a live app is doing dynamic analysis, which is what a human pentester does, at machine speed and without a scoping call.

The Security Ledger is the part to watch

Language models invent things. Put one in a verification loop and the obvious failure mode follows: it reports a flaw as confirmed because the pattern looked familiar, not because the request returned something exploitable.

Detectify's answer is an add-only log written by the tooling itself rather than by the agent. Every test request, payload and result lands in the Security Ledger, and nothing in it can be edited afterwards. An auditor or a compliance reviewer can trace a confirmed verdict back to the actual HTTP exchange that produced it.

Whether that holds up under real adversarial conditions is not something I can confirm from a briefing. Our own measurements at AI Arena are throughput and latency figures for local and cloud models, not security verdict accuracy, so treat the logging claim as a design statement until someone publishes a tamper test.

Four verdicts and one stop sign

Agents must resolve each suspicion into one of four outcomes: confirmed, not confirmed, inconclusive or not testable. That vocabulary is narrower than it looks. Inconclusive and not testable give the model a legitimate way out instead of forcing a binary answer.

Hazardous actions are gated. If the agent decides it needs a denial-of-service style test to settle a question, the action is flagged as unsafe and handed to a human reviewer. Nobody wants an autonomous loop deciding by itself how hard to push a production endpoint.

The triage maths behind the launch

Detectify's argument rests on three numbers. Customers report that about 80% of findings from AI-driven SAST are noise. A study from Fudan University and Tencent Holdings put the false-positive rate of static-analysis alerts at 76%. Developers spend 10 to 20 minutes reviewing each false positive.

Multiply it out. A team facing 100 SAST alerts is spending somewhere between 800 and 1,600 minutes on alerts that lead nowhere, which is 13 to 27 working hours per hundred findings. On a codebase that generates thousands of alerts a month, review becomes the bottleneck and the scanner becomes background noise nobody reads.

The backlog number is worse. Detectify's H2 2026 Cyber Hygiene Index found that 90% of open critical and high-severity vulnerabilities stayed unpatched for more than 90 days. If the queue is not clearing at 90 days, adding a second queue of unverified AI findings is not a fix.

The European angle is mostly about where data sits

Detectify AB is based in Stockholm, and the customer-hosted model keeps source code inside the customer's own perimeter. For EU buyers that removes one item from the procurement checklist: no new processor relationship covering source code, and no transfer question to answer in a data protection impact assessment.

The regulatory backdrop has tightened. Under the Cyber Resilience Act, reporting duties for actively exploited vulnerabilities have applied since September 2026, and NIS2 transposition has pushed vulnerability handling into the obligations of a much wider set of organisations than before. Both reward being able to show when a suspected flaw was tested and what the result was. A timestamped log is closer to that than a screenshot in a ticket.

Europe also has a growing pile of agentic security startups competing for the same budget, and Detectify is not the cheapest way to get a scanner. It is positioning on verification and evidence instead.

What is not public

Pricing. Commercial terms for Agentic Tooling are not listed; access goes through a request or a demo booking, which usually means the number depends on seats, applications in scope, or both. Anyone comparing this against an open-source DAST tool plus a homegrown MCP wrapper is comparing a licence fee against engineering time, and the licence fee is unknown.

Independent accuracy figures are also missing. The 80% noise rate is a vendor-reported customer figure, the 76% false-positive rate comes from academic work on static analysis generally, and neither tells you how often a Detectify agent gets the verdict right. A verification tool whose failure mode is a confident wrong answer is a different risk from a scanner that over-reports.

The rollout is available on request as of 16 November 2026. Detectify has not published a public accuracy benchmark for the four verdicts, and that is the first question a security lead should ask before wiring an agent into a CI/CD pipeline.

Does Detectify get access to our source code?

According to the vendor, no. The AI agents run inside the customer's private environment and connect via Model Context Protocol or a CLI. Detectify states it does not store or access customer codebases.

Can an agent take down a production application while testing?

Potentially hazardous actions, such as denial-of-service style tests, are flagged as unsafe and routed to a human reviewer instead of being executed automatically. Standard test payloads run without that gate.

What does Agentic Tooling cost?

There is no public price list. Access is available on request, so cost depends on a commercial conversation rather than on published per-application or per-seat rates.

Discussion

No comments yet — be the first to share your thoughts.
X

Don't miss out!

Subscribe for the latest news and updates.