Last updated: October 19, 2026
This page describes how Extension Auditor analyzes browser extensions and produces the risk classifications you see on scan pages. Our goal is transparency: developers, researchers, and end users should be able to understand the basis on which we classify an extension and challenge classifications they believe are wrong.
Important. Extension Auditor's risk classifications are automated analytical opinions based on technical signals. They are not statements of fact about the extensions or their developers, and they should not be treated as a substitute for independent security review or professional advice. See our Terms of Service and Privacy Policy for the full disclaimer and lawful basis on which we publish this analysis.
What we analyze
For every extension we score, we collect and analyze the following signals from public sources:
Manifest and permissions
- The full
manifest.jsondeclared by the extension - All requested permissions (e.g.,
tabs,storage,webRequest,cookies) - All host permissions (the URLs the extension may access)
- Content scripts and the URL patterns they match against
- Background scripts, service workers, and their scopes
- Externally connectable origins
- Optional permissions and host permissions
Code analysis
- Static analysis of all bundled JavaScript and HTML/CSS
- Detection of known dangerous code patterns (e.g., dynamic
eval, remote code loading, obfuscated payloads) - Extraction of network endpoints, IP addresses, email addresses, cryptocurrency addresses, and other indicators of compromise
- Comparison against known-malicious code patterns from public threat intelligence
Publisher and listing signals
- Publisher name, contact email, and verification status as published in the Chrome Web Store
- The set of other extensions published by the same developer account
- Listing metadata (description, category, install count, ratings, version history)
- Changes to publisher email or other identity signals over time
Reviews
- Each Chrome Web Store review with text is classified for sentiment (positive, neutral, negative), authenticity (genuine, suspicious, fake) and whether it reads as AI-generated, always against the reviewer's original wording rather than a translation
- Authenticity looks for promotional or off-topic text, disclosed incentives, and boilerplate or keyword-stuffed text. It is judged independently of sentiment, so a harsh or negative review is not treated as fake because it is negative
- Thresholds are set conservatively, because labelling a real person's review as fake is costly. The underlying probabilities are stored so the thresholds can be tuned
Marketplace signals
- Removal status from the Chrome Web Store (active, unlisted, removed)
- Reasons for removal where publicly disclosed (malware, policy violation, voluntary)
- Cross-references to other public databases of malicious or compromised extensions
How risk is classified
Each detected signal is assigned a severity tier and a weight:
| Severity | Weight | Examples |
|---|---|---|
| CRITICAL | 2.0 | Permissions or code patterns associated with credential theft, cryptocurrency wallet exfiltration, or arbitrary remote code execution |
| HIGH | 1.0 | Broad host access combined with cookie or storage access; remote script loading; obfuscated payloads |
| MEDIUM | 0.5 | Single dangerous permission without combination triggers; suspicious code patterns; publisher reputation issues |
| LOW | 0.1 | Permissions with limited risk surface; minor metadata anomalies; informational signals |
The detected signals are aggregated into a numerical score, and that score is mapped to a risk level:
| Risk Level | Aggregate Score | Interpretation |
|---|---|---|
| CRITICAL | 4.0 and above | Multiple severe signals or one signal with very high impact. Extension should not be installed without specific cause. |
| HIGH | 2.0–3.99 | Significant security concerns. Extension should be reviewed carefully before installation. |
| MEDIUM | 1.0–1.99 | Notable security signals. Use with awareness of the analyzed risks. |
| LOW | 0.0–0.99 | Limited security signals identified by automated analysis. |
Certain combinations of permissions trigger a dangerous-combination multiplier that increases the contribution of those permissions to the aggregate score. For example, an extension that combines webRequest, webRequestBlocking, and broad host access has elevated capacity to intercept and modify network traffic; this combination is scored more heavily than the sum of its parts.
The machine-learning score
Separately from the rule-based classification, a machine-learning model produces its own risk score for each analyzed version. It is computed from code features extracted from that version together with listing signals, including review counts and negative-review counts. Reviews from reviewers identified as fake are excluded from those counts, so a publisher cannot move the score by posting fake reviews. The ML score is a distinct signal from the rule-based report and is stored separately.
When an assessment is refreshed
Assessments are tied to a specific version, and several parts refresh on different schedules:
- New versions. Versions that have not yet been analyzed are queued for risk analysis every four hours, so the latest version is assessed shortly after it is collected rather than once at install.
- Monitored extensions. Change detection for monitored extensions runs hourly.
- Reviews. Newly collected reviews are classified hourly. A review whose text is edited arrives as a new review and is classified afresh.
- Machine-learning score. The routine run scores versions that have not been scored yet; re-scoring every version is a deliberate, separate run.
A change to one input, such as a publisher's contact email or a burst of new reviews, does not by itself trigger an immediate re-score of an already-analyzed version. It is reflected the next time the relevant analysis runs.
What we do not classify on
We deliberately exclude the following from our scoring:
- Subjective quality judgments. We do not score extensions on user experience, design quality, or feature completeness.
- Developer demographics. We do not score extensions based on the developer's nationality, organization size, or any protected characteristic.
- Marketplace popularity in the rule-based score. Install count, rating, and review count are recorded as context and do not directly affect the permission, manifest, publisher and code-pattern score described above. A popular extension is not "safer" by virtue of popularity.
- Reviews as a verdict. A review's sentiment never raises or lowers the rule-based score on its own.
- Manual editorial decisions. Risk scores are produced by automated analysis. Human review may be applied to validate findings or address removal requests, but does not by itself raise or lower a score.
Limitations
You should be aware of the following limitations when interpreting Extension Auditor scores:
- Static analysis cannot detect all threats. Code that loads remote payloads at runtime, that activates only under specific conditions, or that is obfuscated using novel techniques may evade detection.
- Extension behavior changes between versions. A score reflects the analyzed version. New versions are re-analyzed, but there is necessarily a window between publication and analysis.
- Permissions are a proxy, not a measurement. Many legitimate extensions request broad permissions for legitimate reasons. The risk classification reflects the capability surface implied by the permissions, not direct evidence of misuse.
- Marketplace metadata can be incorrect. Publisher names, contact information, and verification status are taken from the Chrome Web Store as published. Where this data is wrong, our reports inherit that inaccuracy.
- Legitimate extensions can score HIGH or CRITICAL. A score reflects the analyzed capability and signal surface; it does not by itself prove malicious intent. Consult the detailed scan report and exercise judgment.
Challenging a classification
If you are an extension developer, publisher, or researcher and you believe a classification is unsupported by the underlying signals, you can:
- Submit a factual inaccuracy report at our Removals & Takedowns page. Include the extension ID, the URL of the scan page, and the specific signals you believe are misinterpreted.
- Request publication of a developer response alongside the scan report. We may, at our discretion, publish your response so readers see both perspectives.
- Exercise statutory data-subject rights under GDPR, CCPA, or PDPA where applicable, as described in our Privacy Policy and Removals & Takedowns policies.
We respond to factual inaccuracy reports within 14 days. We do not commit in advance to making any specific change, but every report is reviewed by a human.
Methodology updates
This methodology evolves as we improve our analysis pipeline and incorporate new threat intelligence. Material changes will be reflected in the "Last updated" date at the top of this page. Where a methodology change materially affects existing scan reports, we re-score affected extensions and note the change in our changelog.