Quira Quira Docs
GitHub Star

Threat Protection

Quira's threat protection goes beyond traditional blocklists and heuristics. It introduces an AI Immune System that behaviorally baselines your browser and detects anomalies in real-time, plus adversarial ML defense that protects the browser's own AI from manipulation by web content.

AI Immune System

Inspired by biological immune systems, Quira's AI defense layer continuously monitors browser behavior and autonomous detects threats — even novel zero-day attacks that no blocklist has seen. All security events from the AI Immune System feed into the Context Security Event Bus (CSEB) for programmable rule evaluation and forensic audit.

ComponentWhat it doesExample
Behavioral baseliningLearns normal patterns for each user across resource usage, network patterns, and graph accessDetects a compromised extension suddenly accessing 100x more graph data than normal
Extension monitoringContinuously profiles extension behavior against their declared capabilitiesFlags an extension requesting network access outside its allowlisted domains
Anomaly scoringAssigns real-time threat scores to active processes, tabs, and extensionsQuarantines a tab whose JavaScript is making rapid clipboard reads
Self-healingAutomatically revokes compromised capability tokens and isolates affected SpacesIf an extension's threat score exceeds threshold, its tokens are revoked and the extension is suspended

All analysis is local

The AI Immune System runs entirely on-device. Behavioral models are trained on your patterns locally and never leave the device. No browsing behavior is sent to external services for analysis.

Adversarial ML defense

Quira's AI processes web content to build the Context Graph — which means web pages can attempt to manipulate the AI. This is the browser-specific equivalent of prompt injection, and Quira defends against it with purpose-built countermeasures.

  • Invisible text removal — Detect and strip hidden text (CSS hidden, zero-width characters, white-on-white) designed to pollute AI summaries
  • Unicode normalization — Canonicalize all text before AI processing to prevent homograph-based entity confusion
  • Entity validation — Cross-reference extracted entities against the existing Context Graph to detect implausible injections
  • Embedding anomaly detection — Flag content whose embedding vectors are statistical outliers compared to the page's visible content

AI pipeline 4-layer defense

Every piece of web content that enters Quira's AI pipeline passes through four independent defense layers:

LayerFunctionWhat it catches
L1 — Content SanitizationStrip hidden text, normalize Unicode, remove injected instructionsPrompt injection, invisible SEO spam, adversarial text
L2 — Privilege FramingWrap all web content with explicit "untrusted source" framing before AI processes itContent trying to impersonate system instructions
L3 — Output Schema ValidationAI output is validated against strict schemas — only structured data is acceptedFree-text manipulation, hallucinated entities, action injection
L4 — Capability SeparationThe AI inference process has no capability tokens — it cannot read or write the graph directlyEven if the AI is fully compromised, it cannot exfiltrate data

Living Security

Security in Quira is not a static configuration — it is a continuously adapting system. Living Security introduces the Security Health Score.

  • Security Health Score — A real-time composite score (0-100) visible in the toolbar, reflecting current security posture: extension risk, permission sprawl, outdated blocklists, etc.
  • Adaptive recommendations — When the score drops, Quira suggests specific actions: revoking unused permissions, updating blocklists, removing suspicious extensions
  • Threat memory — The AI Immune System remembers past threat patterns and can recognize variations even after the original threat signature changes

Phishing detection

Quira uses a multi-signal approach to detect phishing pages:

  • URL heuristics — Detects homograph attacks, excessive subdomains, and known phishing URL structures
  • Local blocklist — Regularly updated list of confirmed phishing domains, checked entirely on-device
  • Visual similarity — On-device ML model compares page layouts to known login pages and flags lookalikes
  • Certificate age — Warns when visiting login pages on domains with certificates issued less than 7 days ago

No data leaves your device

Unlike Chrome Safe Browsing or Firefox Phishing Protection, Quira's phishing detection runs entirely on-device. No URLs or page content are sent to external servers for evaluation.

Malware URL blocking

Known malware distribution URLs are blocked at the network layer using a locally maintained blocklist updated via differential sync. Sources include URLhaus, PhishTank, and Quira's own threat intelligence feed.

Certificate validation

CheckDescription
CT log verificationVerifies the certificate appears in at least two Certificate Transparency logs
TOFU pinningTrust-on-first-use: warns if a site's certificate issuer changes unexpectedly
CA reputationFlags certificates from CAs with a history of mis-issuance
Revocation checkOCSP stapling preferred; CRL fallback with local cache

Download scanning

Downloaded files are checked against known-bad hashes before they are saved to disk. Executable files trigger an additional warning dialog showing the file's digital signature status and source domain reputation.

← Previous: Content Filtering Next: Permission System →
Was this page helpful?
Ask AI
Q

Ask anything about the documentation

Ctrl+K to open anytime