Appearance
AI Threat Analysis
GhostAI's AI threat analysis enriches file security findings with AI-powered verdicts. When files are scanned, detected security findings are analyzed by an AI model to determine whether they are truly malicious, suspicious, or benign. This reduces alert fatigue by filtering false positives while surfacing confirmed threats.
Overview
The threat analysis pipeline runs after initial security scanning. Each finding is sent to an AI model with extracted context (e.g., macro code, PDF JavaScript, URLs). The model returns a structured verdict and explanation. Results are cached by content hash for 30 days to avoid redundant analysis.
Key capabilities:
- AI verdicts — Classify findings as malicious, suspicious, or benign
- Adjusted risk levels — AI can downgrade severity for benign findings or upgrade for confirmed threats
- Override indicators — Transparent labeling when heuristics, policy, or low confidence apply
- User feedback — Thumbs up/down on each verdict to improve future analysis
- Download guard — Uses AI-adjusted risk levels to control file downloads
Verdict Badges
Each security finding displays a verdict badge reflecting the AI analysis result:
| Badge | Color | Meaning |
|---|---|---|
| Likely Safe | Green | AI classified the finding as benign (e.g., legitimate macro, print dialog, known-good URL) |
| Suspicious | Yellow | AI could not confidently classify; warrants manual review |
| Malicious | Red | AI confirmed the finding as malicious (obfuscation, exfiltration, shell execution, etc.) |
| Inconclusive | Gray | AI confidence was below threshold; original risk level retained |
| Analyzing | Spinner | Analysis in progress; results will update when complete |
| Unavailable | — | AI analysis is disabled or the LLM is unavailable |
Adjusted Risk Levels
The AI verdict influences the effective risk level shown for each finding:
| AI Verdict | Effect on Risk |
|---|---|
| Malicious | Risk is upgraded to at least HIGH; original severity retained if higher |
| Suspicious | Original risk level unchanged |
| Benign | Risk may be downgraded to LOW or MEDIUM (depending on confidence) |
| Inconclusive | Original risk level retained |
The adjusted risk level is used by the download guard: files with critical findings are blocked; high findings require explicit acknowledgment to download.
Override Indicators
Some findings display an override indicator explaining why the AI verdict or risk level was modified:
| Indicator | Meaning |
|---|---|
| Heuristic | A known-bad pattern (obfuscation, exfiltration, shell execution, exploit) was detected by rules. The finding is treated as malicious without LLM analysis. |
| Policy | Your organization has locked this finding type to a specific severity. Admin policy overrides the AI-adjusted risk. |
| Low Confidence | The AI was not confident enough (below 70%) to classify. The finding remains at its original risk level. |
User Feedback
You can provide feedback on each AI verdict:
- Thumbs up — Agree with the AI verdict
- Thumbs down — Disagree (e.g., you believe a "Likely Safe" finding is actually malicious, or vice versa)
Disagree feedback invalidates the cached verdict for that finding. On the next scan or re-analysis, the finding will be re-evaluated by the AI.
Supported Detectors
AI threat analysis is applied to the following detector types:
| Detector | Description |
|---|---|
| pdf_js | PDF JavaScript/Actions |
| macro | VBA and Excel 4.0 (XLM) macros in Office documents |
| phishing | Suspicious URLs (typosquatting, IDN homograph, shorteners) |
| stego | Potential steganographic hidden data in images |
| csv_formula | Formula injection in CSV/spreadsheets |
| prompt_injection | Prompt injection patterns in file content |
| svg_threat | SVG script injection and XSS payloads |
Skipped (factual): Encrypted archives and zip bombs are not sent to the AI. These are deterministic findings — the system reports them as-is without AI analysis.
Disclaimer
AI threat analysis is advisory. It is designed to reduce false positives and highlight confirmed threats, but it is not a substitute for human review. Always consult your security team when handling files with security findings, especially those marked Malicious or Suspicious.
FAQ
Can I download a file marked Likely Safe?
Yes. The download guard uses the AI-adjusted risk level. If the AI classified a finding as benign and downgraded the risk, you can download the file subject to your organization's policy (e.g., high findings may still require acknowledgment).
Why is a finding still HIGH after the AI says it's benign?
Your organization may have locked that finding type to a minimum severity via admin policy. For example, if pdf_js is locked to high, all PDF JavaScript findings will remain HIGH regardless of the AI verdict. Contact your administrator to change this.
What happens when I disagree with a verdict?
Your feedback is recorded and the cached AI result for that finding is invalidated. The next time the file is scanned or re-analyzed, the AI will evaluate it again. Feedback also helps improve future model tuning.
Related
- File Processing — Security threat inspection and supported formats
- Document Preview — Viewing files with security findings
- Threat Analysis Policy — Admin configuration and controls