PharosOne · Probe Engine
behavioral vulnerability scanner for AI agents · AIUC-1
Run configuration
Target tier
mock — offline, no LLM
model — synthetic agent on a real LLM
bridge — real agent (HTTP endpoint)
Real agent endpoint
OpenAI-compatible chat-completions URL of your/customer agent
API key / Bearer token (if required)
used only for this run, in memory — never written to disk or logs
Attacker model (for adaptive probes)
LLM that drives the adaptive attack; needed only for adaptive probes (uses the same key above)
declare Tools below with dangerous/leak flags to detect abuse of the agent's tools
Provider
Direct (anthropic / openai / google …)
OpenRouter (single key)
Model id
direct: anthropic/claude-opus-4-8 · openai/gpt-4o · google/gemini-2.5-pro
API key
used only for this run, in memory — never written to disk or logs
System prompt of the agent under test
persona of the agent under test (for tier=model); empty = default
Mock rule
by_fingerprint
always
never
Mock ASR %
Industry
Tools (comma-separated; empty = all)
Languages
Seed
Prompt variation
deterministic — seed mutators (offline, free)
llm — paraphrase by a model (key required; on model-tier — the same model)
llm: semantically rewrites prompts (for bridge, set paraphrase_model in the profile/CLI)
Agent description (to tune the attacks)
used by llm variation and the adaptive attack to target the agent's real tools/domain
Variants/probe
Epochs
▶ Start run
Progress
Ready to start. Configure on the left and click "Start".
Probe
Sev
Scenario
Status
ASR
CI
Trials
Taxonomy
Summary
Coverage by control
Control
Title
Status
Probes/Min
ASR
Gap analysis (behaviorally testable, under-covered)
Control
Title
Status
Probes/Min