Switch Trust
Switch Trust is a local-first Python command-line tool. According to its official website, it can scan security risks, misconfigurations, dangerous tool permissions and missing guardrails in the source code of intelligent agents, conduct adversarial evaluations such as prompt injection and jailbreaking on running intelligent agents, and output findings mapped to OWASP ASI as well as reliability scores.
Updated: 2026-10-27
What is Switch Trust?
According to the product page description on the official Switch Trust website, Flint AI has been rebranded to Switch Trust. It positions itself as a free, local-first, framework-agnostic agent command-line tool used to determine the reliability of agents before deployment. The official website describes two commands targeting source code and runtime behavior respectively: `scan` analyzes the source code of Python agents, and `eval` sends adversarial and functional prompts to running agents and assigns scores; the GitHub repository also refers to these two as white-box testing and black-box testing.
Who is it for?
- Developers who need to check the code security and guardrail gaps of Python agents before release.
- Engineering teams that want to conduct prompt injection, jailbreaking, and instruction following tests on running agents.
- AgentOps and security teams that use multiple agent frameworks and want a unified evaluation standard.
- Enterprise engineers who are concerned about the boundaries of source code retention and prefer to run scanning tools locally.
- Technical leaders who need to provide explainable scores and risk inventories for agent launch reviews.
Key features
6Source Code Security Scanning
The official website states that scan can detect misconfigurations, dangerous tool access, missing guardrails, and other code issues, and automatically triages false positives; the GitHub description also mentions that it will output security findings mapped to the OWASP Top 10 with CVSS severity levels.
Runtime Adversarial Evaluation
According to the official website, eval tests behaviors such as factual accuracy, instruction following, prompt injection, and jailbreaking. The evaluation is framework-agnostic and provides a reliability score ranging from 0.0 to 1.0.
Multi-framework Support
The official website lists frameworks such as Google ADK, Anthropic SDK, OpenAI SDK, LangGraph, CrewAI, AutoGen, HuggingFace Transformers, and smolagents; it runs intelligent agents in eval tests, so it can also cover frameworks outside the list.
Local Execution Boundaries
The official website states that the CLI runs locally with no backend, telemetry server, or account system—source code remains on the local machine; only data sent to the configured LLM providers leaves the device.
Installation & Dependencies
The official website states that the CLI is free, requiring Python 3.11 or higher and an LLM API key. It supports OpenAI, Anthropic, and Gemini. GitHub also lists the installation requirements for scanning and evaluation.
Scoring Criteria
The official website states that each AI agent will receive a reliability score, with findings mapped to OWASP ASI01 to ASI10 and scored using CVSS v4.
How to use Switch Trust
- First run `scan` to address code-level issues, then launch the agent and run `eval`.
- the two cover white-box and black-box risks and cannot replace each other.
- Sort scan findings by OWASP classification and severity, and prioritize addressing issues that impact production security, such as dangerous tool permissions, missing authentication, and infinite execution loops.
- eval requires the agent to be accessible and the model to be configured.
- First, calibrate the workflow using official examples, then integrate your own test sets and business rules.
- The official website states that the source code remains on the local machine, but scanning and scoring will still call the selected LLM provider.
- key management, the provider's data processing policies, and corporate data boundaries should be evaluated together.
- GitHub mentions that anonymous telemetry requires consent upon initialization and will not be sent without explicit consent.
- in enterprise environments, this can be uniformly disabled before deployment with clear internal policies documented.
Traffic & popularity
Data month Sep 2026
Switch Trust receives about 960 visits per month.
Monthly visits
3 monthsEstimated from third-party traffic data.
In these collections
Similar tools
AntigmaAntigma is a terminal coding agent and agent runtime: a single command to install a single-file program, switch between cloud or local models in a session, built with Rust and with no external runtime dependencies.
SourcerySourcery is a team-oriented AI code review tool. After integrating with GitHub and GitLab, it automatically reviews every pull request, detects defects, scans for security vulnerabilities, and provides one-click repair suggestions.