Platform
Find it before someone else does
Every agent you ship is reachable by someone who will try things your test suite never considered. SAF3AI runs those attacks first — on a schedule and in your pipeline — and returns findings in a format your existing review process already handles.
Capabilities
Key Features
Attack your own agents
Automated adversarial testing against the agents and assistants you actually run — prompt injection, jailbreak, data extraction, tool abuse and excessive agency, run as an attacker would.
Judged by a model, not a regex
Results are assessed by an LLM-as-judge with a scoring threshold, so a refusal is scored as a refusal and a partial compliance is not counted as a clean pass.
A gate in your pipeline
The CLI returns standard exit codes against a severity threshold, so a build that introduces a regression in agent safety fails rather than ships. A GitHub Action wraps it for convenience.
SARIF findings
Results come back as SARIF 2.1.0, which lands them in GitHub code scanning, your IDE and whatever triage flow you already run — instead of another console to check.
Scheduled, not just on demand
Agents change, models get updated, prompts get edited. Scheduled scans catch the regression that arrives when nobody touched the safety logic on purpose.
Findings become graph entities
A red-team finding attaches to the agent it concerns, so posture, runtime detection and adversarial testing are three views of one entity rather than three separate reports.
In your pipeline
One engine, three places
The CLI is the single implementation. The IDE extension and the CI action both invoke it, so a finding in your editor and a finding in the pipeline come from identical logic — not from two codebases that drift apart.
Terminal
Run a scan against an agent or an endpoint, watch it progress, and read the findings where you are already working.
IDE
The VS Code extension renders findings as diagnostics inline, so an issue surfaces next to the code that caused it.
CI
A composite GitHub Action, or any pipeline that can run a binary. Exit codes gate the build against your chosen severity threshold.
ci-pipeline # Fail the build on any high-severity finding
saf3ai redteam scan --agent checkout-assistant --fail-on high
# Scan AI-generated code before it merges
saf3ai sast scan ./services/agent --fail-on medium -o sarif > results.sarif
# Exit codes: 0 clean · 1 findings at or above threshold
# 2 usage or auth · 3 runtime errorGet started with Red Teaming
See how Saf3AI can help secure your AI agents.