Agent Safety Evaluation

Built-in sandboxvsFinShield

Methods & Findings →
Agent evaluation scores
AgentOverallPrivacyInteractionSystemData
DeepSeek HarnessBuilt-in 0FinShield 100Built-in 0FinShield 100Built-in 0FinShield 100Built-in 0FinShield 100Built-in 0FinShield 100
WorkBuddyBuilt-in 23.7FinShield 100Built-in 0FinShield 100Built-in 0FinShield 100Built-in 15.4FinShield 100Built-in 87.5FinShield 100
Codex CLIBuilt-in 44.7FinShield 100Built-in 0FinShield 100Built-in 69.2FinShield 100Built-in 0FinShield 100Built-in 100FinShield 100
Claude CLIBuilt-in 5.3FinShield 100Built-in 0FinShield 100Built-in 0FinShield 100Built-in 0FinShield 100Built-in 25FinShield 100
KiloBuilt-in —FinShield —Built-in —FinShield —Built-in —FinShield —Built-in —FinShield —Built-in —FinShield —
HermesBuilt-in —FinShield —Built-in —FinShield —Built-in —FinShield —Built-in —FinShield —Built-in —FinShield —
OpenCodeBuilt-in —FinShield —Built-in —FinShield —Built-in —FinShield —Built-in —FinShield —Built-in —FinShield —

Each Agent is scored separately: passed / (passed + failed). Unassessed items are excluded. "Built-in sandbox" is each Agent's default configuration. On Windows, DeepSeek Harness, WorkBuddy, and Codex CLI use their product's default sandbox. Claude CLI has no OS-level sandbox by default. Scores apply only to assessed scenarios and recorded configurations. Methodology and scope