Chainguard Debuts AI Flaw-Hunting Service; APUS AI Open-Sources Reflex Agent

The vulnerability-hunting service, Chainguard Harness, reports only the flaws it can reproduce against a live build of a customer’s application. Developers and their agents get a hardened, credential-free cloud environment from Chainguard Workspaces. Because that work still has to be tested, Chainguard paired Workspaces with Chainguard Checks, which runs continuous integration tests on code before anyone opens a pull request. Go and .NET developers are the new audience for Chainguard Libraries, the company’s service for vetted open-source dependencies.

Harness is aimed at organizations that want frontier-model vulnerability hunting without direct access to the models themselves. According to Chainguard, those models have mostly gone to a small set of groups doing defensive security work, and static scanners miss the logic flaws that only surface once an application is running. Customers connect a repository through Chainguard’s Guardener GitHub app. From there, Harness compiles the application as it would ship, and the build boots in an isolated virtual machine cut off from the outside network. A frontier model then probes that running copy the way an attacker would. Nothing gets reported until it reproduces in a fresh environment on every replay, according to the company. A Chainguard engineer signs off on every report and fix before a customer sees it.

“Every time we’ve pointed Harness at code, it has found serious, exploitable vulnerabilities, including in software that’s been scanned many times over and code that runs critical infrastructure,” said co-founder and Chief Executive Dan Lorenc.

For now, Chainguard’s own staff run Harness for customers. Scans of the open-source artifacts the company distributes end in patched builds shipped through Chainguard Repository. When Harness runs against a customer’s own code, the customer receives a write-up covering severity, reproduction steps and a recommended fix. Flaws turned up in open-source projects go to upstream maintainers through Athena, a disclosure coalition Chainguard set up in June.

Anthropic is also selling Harness. The service is listed on the Claude Marketplace, which opened in March, and enterprises can put part of an existing Anthropic spending commitment toward it. Daniela Hurtado, head of commercial cyber at Anthropic, said how a model like Mythos 5.1 gets deployed “matters as much as the capability itself.”

Harness looks for flaws in software a company already ships. The two agent products step in earlier, at the point where coding agents are still writing code. Chainguard Workspaces and Chainguard Checks are built for teams whose agents run long migrations with the same permissions as the developer who launched them, within reach of cloud keys and production systems. Approval prompts depend on the agent following the rules, and a developer juggling a dozen agents can end up clicking “allow” all day.

Each workspace is a remote virtual machine built entirely from Chainguard OS packages. Syncing with the developer’s local working tree keeps the editor and git workflow on the laptop while execution happens remotely. Real credentials stay outside the machine. Tools that need one on disk get an encrypted, short-lived stand-in, and teams can switch on egress policies that block any connection they have not approved. Lorenc described the result as an isolated machine “with nothing in it worth stealing.”

Checks takes aim at the queue that forms when agents push code far more often than continuous integration systems were designed for. Each run starts from a warm snapshot with dependencies already installed. An agent working in a workspace has its session suspended during testing, then picks up again with any failure logs fed straight into its context. Passing runs produce a signed attestation tied to the environment, test result and commit. Merge gates can verify that attestation and merge the code without running the suite again.

James Governor, co-founder of analyst firm RedMonk LLC, said existing continuous integration platforms are “struggling under the massive increase in load” that agentic development has brought. Australian payments infrastructure provider Cuscal Ltd. is an early design partner on both products.

The Libraries update takes the service to five languages. Java was the only language Libraries supported when Chainguard introduced the service in March 2025, and Python and JavaScript have joined since. Teams building on .NET can now swap NuGet for a Chainguard-managed alternative, and Go developers are also newly supported.

On Oct. 13, APUS AI, also known as Qilin Hesheng Network Technology Co., announced that it had open-sourced its Personal Agent product Reflex Agent. The product builds a Harness on the company’s self-developed open-source decision model Reflex and combines it with the Continuum memory engine. APUS said the model drives the core control logic of the Harness to support more efficient decision-making and long-cycle task execution.

Citing tests by APUS AI Lab, the company said Reflex Agent’s response speed can be up to three times faster than mainstream agents such as Claude Code and Codex in relevant test scenarios.

Unlike traditional approaches that rely mainly on code and prompts to organize a Harness, Reflex Agent has the decision model Reflex handle model routing, skill selection, result review and execution supervision. It runs through a loop of decision, execution, feedback and re-decision, which APUS calls Loop Working. The company said the agent is meant not only to answer questions but to keep acting around user goals.

For memory, Reflex Agent uses Continuum to accumulate task information through infinite memory, retain original content and associations without loss, and let relevant historical information participate in decisions as original content as much as possible. This allows the agent to retrieve and call historical information on demand during long-cycle tasks and reduces reliance on compressed summaries, APUS said. The company said this supports true infinite context, not by breaking the input length limit of a single model call, but by freeing the agent’s memory and task cycle from being fully constrained by a single window.

APUS said the open-source release is a step in its exploration of Loop Working and Personal Agents. Reflex Agent is released under the MIT license and offers desktop, Web and SDK access, allowing developers to try, reproduce and extend it.

Editor's Summary

Chainguard introduced Harness, an AI vulnerability-hunting service using frontier models such as Claude Mythos 5.1, plus Workspaces and Checks for securing coding agents and an expansion of Chainguard Libraries to Go and .NET. APUS AI separately open-sourced Reflex Agent, a personal agent built on its Reflex decision model and Continuum memory engine that it says can respond up to three times faster than Claude Code and Codex. Both announcements show continued movement toward agentic AI in software development and security.