Topsec launches an AI penetration-testing agent that refuses unverified findings

Topsec (002212) unveiled a new AI-driven automated penetration-testing system on November 3 at the "Light of the Internet" Expo, part of the 2026 World Internet Conference in Wuzhen. The product is described as a security-controllable penetration-testing agent built on the company's self-developed RHP architecture; product director Liu Desen presented it at the event's digital-innovation launch session.

Traditional penetration testing is bounded by signature libraries and preset rules, leaving risks outside those rules in a blind spot. Topsec argues that real threats increasingly take the form of logic flaws, long chains and combined vulnerabilities: a single point may not look critical, but chained together they can lead to lost privileges. The new system therefore emphasizes inference over pattern matching.

The RHP architecture coordinates multiple agents with autonomous planning, reasoning and judgment. Starting from asset mapping, the system performs data-flow analysis, maps permission relationships and reasons about multi-step reachable paths, trying to understand business logic in context and connect isolated risks into a complete attack chain. That shifts penetration testing from comparing known signatures to inferring unknown risk.

False-positive control is the product's most explicit principle: unverified vulnerabilities are not allowed into the report. Every finding must come with a reproducible proof of concept, exploit and the raw request and response, with one-click retesting so the result can prove itself on the spot. Topsec also keeps evidence along the attack path, from entry point to trigger, so remediation teams can see how a system was actually broken.

Execution is automated end to end: task planning, action execution and result assessment are handled by agents, and the system can run around the clock. Traditional testing depends on scarce experts, which limits scale and often restricts critical assets to periodic checks, leaving gaps that attackers exploit. Automation is meant to move experts away from planning and verification and make continuous testing practical.

The capability layer draws on more than 30 years of offensive and defensive work, packaged as in-house tools, SKILLS plugins and knowledge bases that agents can call. Topsec's framing is blunt: bolting a scanner onto a general-purpose model produces at best a scanner that writes reports. The agent is the brain, security tools are the hands and feet, and the real asset is the attack playbooks refined over years.

Guardrails are treated as part of the product, not an afterthought. Authorization rules constrain targets, permissions and scope; a human-in-the-loop step requires approval for high-risk actions; and an ontology-based blackboard governs agent behavior. The company's line is that the smarter the agent, the harder the guardrails must be.

Against a backdrop of widening AI-driven attack-and-defense asymmetry, the signal matters more than any single specification: a security vendor is trying to rebuild penetration testing around agents and a verifiable evidence chain, rather than tools and scarce experts. The real test will be at customer sites—unknown-vulnerability discovery, false-positive rates, stability in complex business environments and the approval experience for high-risk actions.