Prompt Injection Dataset Scanner v2

Scan buyer-provided text or dataset rows for explainable prompt-injection patterns.

What this Actor does

Return severity, matched signals, evidence snippets, field paths, and review recommendations.

• Run prompt-injection regression checks.
• Locate suspicious instruction and exfiltration language.
• Export explainable findings for human review.

Who this is for

AI security teams, Dataset curators, Agent developers.

Input

• Demo mode is the safe default and emits one deterministic fixture.
• Live mode processes at most maxItems configured inputs.
• records: buyer-provided JSON rows.
• sourceDatasetId: optional existing Apify dataset.
• textFields: optional field paths to scan.

Output

The default dataset uses the Actor-specific schema in .actor/DATASET_SCHEMA.json. Key fields include recordIndex, scannedFields, severity, riskScore, matchedSignals, evidenceSnippets, recommendedAction, plus exact source, status, delta, confidence, warning, and fixture/live labels.

Data sources

• APIFY-DATASETS

This is the registered catalog source ID. Live inputs may also include buyer-provided inline records; buyer inputs are not registered source IDs.

Demo run

Run examples/first-run.json. The row in examples/sample-output.json is labeled sourceStatus: fixture, isDemo: true, and is not live evidence.

Live-run behavior

Live mode processes only the configured Actor-specific inputs. Failures are returned as diagnostics or explicit errors; the runtime does not silently substitute unrelated source families.

Pricing

The runtime reads the effective Apify pay-per-event map after Actor.init() and writes each buyer-visible dataset record through the verified result-delivery event. It stops with an explicit configuration-drift error instead of emitting unmetered output. This package never defines or changes provider pricing.

Identity and visibility safety

This package targets only zentrafoundry/prompt-injection-dataset-scanner-v2 with Actor ID p0oDTPPu7WV13f0jj. Any future release must find exactly one authenticated inventory match and abort on identity or visibility drift. Creating, cloning, upserting, duplicating, drafting, or privatizing the Actor is forbidden.

Limitations

• Heuristic findings can include false positives and false negatives.
• A clean result is not a guarantee that content is safe.
• Local tests and fixtures do not prove an Apify build, live run, Store content update, or Store readback.

Responsible use

Use only data you are authorized to process. Respect source terms, robots policies, privacy rules, and applicable law. Never place credentials, cookies, tokens, or private account data in URLs or logs.

Support

When reporting a problem, include the run ID, sanitized input shape, expected result, and observed result. Do not include secrets or private data.