Running an Attack

Overview of the attack flow

Projects tab → Select project → Select target → Configure → Start Attack
     ↓
Discovery Phase  (UI detection, ~10–60s, runs once)
     ↓
Recon Phase      (optional, probes AI capabilities)
     ↓
Attack Loop      (per goal: strategy → prompt → response → Judge score)
     ↓
Results uploaded to Web Console

Step 1 — Select a project and target

Open the Projects tab. On the left, select your project. In the center panel, select the target you want to attack.

Step 2 — Configure and start

Click Start Attack (or the play button next to the target).

In the configuration panel:

Option	Description
Attack mode	Web UI — browser automation. API (Auto-Detect) — captures the target's API from network traffic. Both — tries API first, falls back to Web UI.
Goals	Select which goals to attack. Goals are synced from your Web Console project.
Max turns	Maximum prompt attempts per goal before giving up (default: 10).
Force Recon	Run a reconnaissance phase first to probe the target's capabilities and system prompt. Adds ~2 min.

Click Start. The Adapter opens a browser window for the target.

Step 3 — Discovery phase

The Adapter automatically finds the chat UI:

• With a known target preset — detection completes in ~10 seconds
• Without a preset (generic) — Vision + DOM analysis runs, takes ~30–60 seconds

You'll see the discovery status in the Live Activity panel at the bottom of the screen. If discovery fails after 3 attempts, see the troubleshooting section below.

Step 4 — Watch the Live Activity panel

During the attack, Live Activity shows each turn in real time:

• The strategy name (e.g. roleplay, indirect, multilingual-ja)
• The prompt sent
• The AI's response (truncated)
• The Judge score

A turn marked Breached (score ≥ 0.7) means the attack succeeded for that goal. The Adapter moves on to the next goal automatically.

Step 5 — Review results in the Dashboard

When the attack finishes, results are uploaded to the Web Console automatically.

In the Adapter's Dashboard tab:

• Overview — total ASR, breach count, goal-by-goal summary
• History — individual assessment list with timestamps

For full analysis including trace details and report generation, open the assessment in the Web Console.

---

Troubleshooting common errors

Error	Cause	Fix
"Chat could not be loaded"	Stale cached session (common with Gemini)	Sessions tab → delete the target session → re-create it
401 Unauthorized	API key expired or revoked	Web Console → Settings → API Keys → create a new key, re-login in Adapter
"Insufficient Credits" (402)	Out of credits	Web Console → Billing → Top Up, then restart the attack
Discovery fails after 3 attempts	Target UI changed or Shadow DOM/iframe not detected	Try switching attack mode to API (Auto-Detect); if unavailable, contact support
Response shows "thinking..." forever	Loading text pattern not recognized	Known issue with some streaming UIs — the Adapter retries automatically
"Target not in Allowlist"	Target URL not approved	Settings → Allowlist → submit the URL and wait for approval
Attack stops mid-run	Credits ran out during the run	Top up and re-run — completed goals are saved, only unfinished goals need re-running
Network Capture failed	Target API is encrypted or obfuscated	Switch to Web UI mode instead of API (Auto-Detect)

Security notes

• Target session cookies are stored only on your local machine — never sent to Stinger's servers
• The Adapter enforces Allowlist restrictions — it will not attack unapproved URLs
• Keep your API key secure. Rotate it at Settings → API Keys if you suspect it was exposed