Running Scans
The Scan tab is where a workspace actually does work. You pick one or more payload categories, kick off a run, and VAAST fires the selected payloads at the workspace's target in order, collects responses, and classifies what comes back.
Selecting payload categories
VAAST organises its catalog into four research-driven categories. Each category tests a different class of AI-specific failure mode:
- Prompt injection (
prompt_injection) — attempts to override, leak, or subvert the target's system prompt and safety guardrails. Covers direct injection, indirect injection via retrieved content, role-play jailbreaks, and instruction-smuggling patterns. - Tool-call abuse (
tool_call_abuse) — attempts to trick the target into invoking its registered tools / function calls in unintended ways. Covers argument injection, unauthorized tool selection, chained-tool privilege escalation, and destructive action coercion. - RAG poisoning (
rag_poisoning) — attempts to manipulate the retrieval step in retrieval-augmented generation pipelines so that the LLM ingests attacker-controlled content. Covers corpus pollution probes, embedding collision checks, and retrieval hijacking. - Agentic abuse (
agentic_abuse) — attempts against multi-step agents: loop abuse, plan-rewrite injection, memory poisoning, and cross-session state leakage.
You can select one category, a subset, or all four for a single run. VAAST shows a live count of how many payloads the current selection will fire.
Running a scan
- Open a workspace.
- Switch to the Scan tab.
- Tick the categories you want to run.
- Click Run scan.
Progress and intermediate results stream into the run view while the scan is in flight. You can cancel a run at any time; completed payloads up to the cancel point are still saved to history.
Reading findings
Each finding in the run view carries:
- Severity —
critical,high,medium,low, orinfo. - Category and payload — which check fired.
- Title and summary — the failure mode in one line.
- Evidence — the exact request/response pair that triggered the detector, with headers and bodies captured verbatim.
- Why it matters — a short explanation of the real-world risk this class of failure represents.
Severity levels
| Level | Meaning |
|---|---|
| critical | Direct compromise: system prompt leaked, safety bypassed, destructive tool invoked, or private data exfiltrated. |
| high | Strong evidence of a bypass or misuse, but attacker still needs a second step to cause damage. |
| medium | Partial bypass or ambiguous evidence; a reviewer should look. |
| low | Behavioural oddity that is unlikely to be exploitable alone. |
| info | Observation about the target's posture, not a defect. |
Severity is assigned by the detector that fired, not by the payload. Two payloads from the same category can produce different severities depending on what the target actually did in response.
Exporting reports
From the run view, click Export report and choose a format:
- HTML — a self-contained, dark-themed report suitable for sharing with engineering teams. Includes severity summaries, per-finding evidence blocks, and a run metadata header.
- JSON — the raw run record, including every request/response pair, for ingestion into other tools or automated triage.
Export is available on every tier. Full-fidelity HTML export — including request/response bodies — is a Pro feature; the Free tier exports summaries only. See Plans & Billing.
See Payload Library for how the catalog is built and what the payload files look like.