subagentic.ai RSS X YouTube GitHub
AI-generated illustration

News

Detectify lets a team's own agents retest suspected flaws on live apps

Detectify’s Agentic Tooling lets a team’s own agents retest suspected flaws on live apps, with a ledger they cannot edit.

detectifyappsecmcpagents

On November 16, 2026, Detectify launched Agentic Tooling so a team's own AI agent can test a suspected flaw against a running application, while Detectify records the evidence in a ledger the agent cannot edit.

The interesting part is not another hosted model. Detectify does not host or run the agent. The product is a verification loop: suspected issues come in, the customer's agent tests each one, and every verdict is supposed to come with proof. Product marketing manager Sanjidah Ahmed's post calls that an automated verification factory. SiliconANGLE's Duncan Riley, writing the same day, frames it for DevSecOps teams that are not sure they can trust what their agents report.

One suspected issue, tested live

Agentic Tooling is for security teams that already run agents, or are ready to, and want evidence instead of a triage pile. Submissions can be static-analysis findings, bug bounty reports, or issues an internal red team reports itself. SiliconANGLE also lists findings from Detectify's own attack surface data. Each submission gets its own ID and becomes a ticket: one focused task.

The agent runs in the customer's environment and calls Detectify through an MCP server. Ahmed's post names Claude Code and OpenCode as examples, and says teams keep their own model and setup. SiliconANGLE adds that the tools can also connect over a command-line interface, operate only on data the customer provides, and are said to work with whichever agent the customer already uses.

Detectify uses what it calls loop engineering: iterative workflows instead of a person prompting a model at every step. SiliconANGLE describes a trigger, such as a continuous integration run, starting each loop. The agent reads the relevant code, then tests the live application from the outside with Detectify's payload-based assessments. SiliconANGLE reports that, through Agentic Tooling, agents can tap the testing engines Detectify has built over more than a decade. Included are payload assessments sourced from its ethical hacker community, attack surface data from Surface Monitoring, and Alfred, an AI system the company launched in March 2025 to automate tests for known vulnerabilities.

The question is narrow. Detectify's FAQ says this is only partly AI pentesting. The agent runs real attacks against a live system, but it answers whether this specific suspected issue is real, one issue at a time. The result is a verdict with evidence, not a pentest report. It is also not continuous DAST. Detectify's Application Scanning tests running apps for whatever it can find. Agentic Tooling verifies a named suspicion, using the same testing engine. As new tickets arrive, that repeating loop is what Detectify calls a verification factory.

A ledger the agent cannot edit

The Security Ledger is the record Detectify's Agentic Tooling maintains. It captures the assessments run, the agent's actions, and most of the investigation traffic, including tool calls, requests, and payloads, and it ties those entries back to the original tickets. The agent cannot edit it. SiliconANGLE describes the ledger as append-only and written by Detectify rather than the agent.

Detectify's distinction is explicit: an agent's reasoning is probabilistic, but the record of what was actually tested is deterministic. If an agent reports a vulnerability and the ledger shows no test behind it, the finding is rejected. Engineers can check the work against their own standards instead of trusting the write-up.

That rule is aimed at a failure mode Detectify is already describing. SiliconANGLE reports that Detectify customers say roughly 80% of findings from their AI-driven static application security testing tools are noise. The company also warns that an agent under pressure to answer can produce a convincing report about a test it never ran. The same article cites Detectify's H2 2026 Cyber Hygiene Index, which found that nine in 10 open critical and high-severity vulnerabilities had stayed exposed for more than 90 days, and a study by researchers at Fudan University and Tencent in which 76% of the static-analysis alerts examined were false positives, each taking developers 10 to 20 minutes to review.

Co-founder and chief executive Rickard Carlsson told SiliconANGLE: "Ask an agent to write your code, then ask the same agent whether it's secure. If it says yes, you don't actually know." He described the aim as "building proof that an agent can't fabricate."

Not every ticket is left to the agent. SiliconANGLE reports that a ticket that could cause harm, such as a denial-of-service test, is marked unsafe for a person to decide on.

Not the May data server

This is a different product from the Detectify MCP Server. A May 26, 2026 item on Detectify's blog describes that server as delivering vulnerability data and attack-surface insights into AI workflows, and SiliconANGLE says it was announced in May so agents could reach Detectify's findings. Ahmed's FAQ draws the line: the MCP Server gives an agent Detectify's data. Agentic Tooling gives it testing tools, a work queue of tickets, and an evidence record in the Security Ledger that it cannot edit.

Source code stays on the customer side. Detectify says the tooling needs access to source code inside the customer's environment so an issue can be traced to where it is exploited in the running app. Detectify says it does not access or store the codebase. It receives the final investigation data, ledgers, ticket contents, and request metadata. SAST exports can include code snippets. Teams that would rather not share those can describe the issues in plain text instead.

What to do next

If you already have an agent that speaks MCP, read Ahmed's post for the FAQ on source-code handling and the split from the May data server, then book the demo if you want the workflow tried on your own findings. That page also links to a trial signup. SiliconANGLE reports the tools are available on request starting November 16, and it is the place to read Carlsson on proof an agent cannot fabricate.

Sources