QAVision · Security and safety
Guardrails in code, not in prompts.
Telling an AI agent to behave is not a control. Every safeguard below is enforced by the system around the agents, and each one is tested.
Sandboxed execution
Code an agent writes runs in a throwaway container: only the working copy mounted, no secrets in its environment, all privileges dropped, CPU and memory capped. If the sandbox isn't available, the work stops rather than running unprotected.
Scoped file access
Agents can read only their working directory. Secret files are masked even inside it, and we prove it with planted canary files on every change.
Results verified by code
"The agent says the tests passed" is not evidence. The pipeline records every test run and labels each result verified, failing or not run. All-skipped counts as not run.
A person approves every write
Agents open pull requests and draft tickets. People merge, publish and close. Only approved accounts can assign work to an agent.
A second opinion from another model
A reviewer from a different AI model family checks the main model's answers and changes, so one model's blind spots don't go unchallenged.
Audit, pause and cost caps
Every action goes to a hash-chained audit log that shows if entries are removed. A pause switch stops all agents at once, and spending is capped per request and per ticket.
Secrets stay out of reach
CI jobs that hold credentials run only reviewed code from your main branch, never code from a pull request, including one an agent opened.
Your AI providers, your rules
AI calls go only to providers you approve, with personal data redacted before a prompt leaves your environment and every call logged.
Bring your security team
We walk through the architecture, the data flows and the tests behind each control, and answer your review questions in writing.