From Chrome to AI Agents, “Sandbox” Is Becoming a Sales Pitch
How Chrome earned users’ trust in sandboxing, and what Windows tests of three AI agents show about device access, data protection, and approval prompts.
How we test Agent safety, design cases and interpret the evidence.
How Chrome earned users’ trust in sandboxing, and what Windows tests of three AI agents show about device access, data protection, and approval prompts.
One question at a time: the design and reasoning behind the cases.
Why evaluate Agent safety? Start with computer permissions and user risks, then explore the evaluation goals, methods and findings.
Pop-ups, black overlays and changed mouse settings: interaction safety tests with four Agents, and FinShield's results and protection mechanisms.
From prompt injection risks to camera, microphone, screenshot and screen recording tests with four Agents. See the prompts and recordings.
Separate program validity, Agent execution and protection outcomes, and explain what each case design can measure.
Distinguish initial blocks, warnings, approval decisions and final effects, and explain why an approval alone cannot establish a safety result.
Use file access and parent-directory operations to explain why targets, operation paths and resource boundaries need separate checks.
Connect attempted operations, independent observations, warning captures and unassessed states into a result that readers can verify.