74.6% of Organizations Report Rising Code Risks From AI Assistants
As agentic coding generates more software than human teams can review, 74.6% of 309 surveyed respondents agree that AI coding assistants increase the risk of developers merging code without fully understanding its security implications. That figure stems from a July 2026 IDC DevSecOps and software supply chain security survey. Software delivery vendors are responding with infrastructure controls designed to isolate automated coding agents and block unverified changes before they reach production environments.
Harness launched its freshly acquired Cosmos software factory on October 8. Chainguard followed on October 14 with Workspaces and Checks. Both platforms utilize isolated virtual machines to run coding agents. Cosmos coordinates multiple parallel agent runtimes, while Chainguard Checks lets agents run tests and correct problems during development without waiting for a traditional continuous integration pipeline.
Isolated MicroVMs Replace Unprotected CI Runners
Cosmos and Chainguard Workspaces spin up ephemeral microVMs to give agents a separate, disposable execution environment. This architecture isolates untrusted code while maintaining a lighter infrastructure footprint than traditional long-running virtual machines. Suresh Gangula, a DevOps engineer at Nike, noted that this approach treats an agent like an untrusted CI runner rather than a developer with unrestricted access.
A disposable VM gives each task a separate execution boundary and reduces the blast radius if an agent, dependency or credential is compromised,
Gangula said. Traditional CI runners often hold long-lived secrets alongside untrusted code, turning pipelines into attractive supply chain targets. Chandrasekhar Katru, a software engineer at Wells Fargo, explained that microVMs provide independent kernels that prevent a compromised container from breaching the host machine.
Network Controls and Automated Verification Guardrails
Beyond workload isolation, Chainguard Workspaces and Harness downstream worker agents implement network controls to restrict where agents can send data. Katie Norton, an analyst at IDC, observed that Chainguard keeps secrets out of the environment entirely and allows teams to set egress as tightly as each task requires. If a narrow, repetitive agent loop attempts an unauthorized outbound connection, the environment can be torn down and the task retried.
In both Cosmos and Checks, agents respond to failed tests and correct their work without human intervention. Sam Katzen, director of product marketing at Chainguard, described Checks as a replacement for GitHub Actions. However, Norton cautioned that replacing established CI platforms remains difficult because pipelines carry years of organizational logic.
Autonomous Software Factories Face Practical Oversight Limits
Routine tasks such as fixing bugs, remediating vulnerabilities, and implementing small feature requests account for 60% to 80% of coding work, according to Jyoti Bansal, co-founder and CEO of Harness. Bansal projected that Cosmos could automate 40% to 50% of software engineering tasks autonomously. Practitioners remain skeptical, citing incomplete tests, incorrect requirements, and high-risk production changes that still require human judgment.
I would not say an autonomous software factory has been established yet,
Katru said. The trick is getting the product reviewed and merged.
IDC survey data supports that caution: among organizations using AI coding assistants, only 28.0% require peer review of AI-assisted changes, and only 23.3% log or audit assistant usage, leaving a wide visibility gap as automated software factories scale.
Related reading