HashJack: The Stealthy Exploit Weaponizing Trusted Sites Against AI Browser Users
Share this article
HashJack: The Stealthy Exploit Weaponizing Trusted Sites Against AI Browser Users
 injects the fragment into the LLM's context.
Malicious Execution: The AI delivers tainted responses, from scam links to background data leaks in agentic models.
"HashJack represents a major shift in the AI threat landscape, exploiting two design flaws: LLMs' susceptibility to prompt injection and AI browsers' decision to automatically include full URLs, including fragments, in an AI assistant's context window," Cato CTRL researchers explained in their detailed report.

Why HashJack Evades Detection and Amplifies Risks
This attack thrives on subtlety. URL fragments remain client-side, never transmitted to servers, rendering network-based defenses like firewalls or IDS ineffective. It chains trust: users trust the site, the browser, and thus the AI's output.
For developers and security engineers, HashJack highlights critical gaps in LLM architectures:
No Fragment Sanitization: AI browsers ingest full URLs indiscriminately.
Agentic Escalation: Browsers like Comet can autonomously exfiltrate data.
Industry implications are profound. As AI agents integrate deeper into workflows -- from code completion to financial analysis -- unsanitized web context becomes a poisoned vector, demanding robust input validation and context filtering.
Attack Scenarios: From Phishing to Silent Exfiltration
Cato demonstrated HashJack's versatility:
Phishing Augmentation: On a support forum, the AI appends fraudulent contact details to responses.
Stock Manipulation Lies: Querying market news on a finance site yields fabricated surges (e.g., "Company X up 35% this week").
Data Harvesting: In Comet, a banking query like "Loan eligibility?" triggers silent POSTs of transaction data to attacker servers.
graph LR
A[Legitimate URL + #Malicious Fragment] --> B[User Visits Site]
B --> C[AI Browser Query]
C --> D[Fragment Injected to LLM Context]
D --> E[AI Outputs Phishing/Data Theft]
style A fill:#ffcccc
style E fill:#ffcccc
Vendor Disclosures and Patch Status
Reported in August 2025:
| Vendor | Product | Status | Notes |
|---|---|---|---|
| Microsoft | Copilot for Edge | Fixed (Oct 27) | Defense-in-depth against variants |
| Perplexity | Comet | Fixed (Nov 18) | Critical severity via Bugcrowd |
| Gemini for Chrome | Won't Fix | Low severity (S3/S4), intended behavior | |
| Anthropic | Claude for Chrome | Resisted | N/A |
| OpenAI | Atlas | Resisted | N/A |
ZDNet reached out to Google for further comment.
Navigating the Evolving AI Threat Landscape
HashJack isn't just a browser bug -- it's a harbinger of sophisticated, low-friction attacks on LLM ecosystems. Developers integrating AI assistants must implement fragment stripping, prompt guards, and behavioral monitoring. Users, meanwhile, should query AI browsers cautiously after clicking shared links, especially on sensitive domains.
As AI browsers evolve toward greater autonomy, so do the stakes. Cato CTRL's revelation compels the industry to rethink URL handling in intelligent interfaces, ensuring that convenience doesn't come at the cost of security. In this new frontier, vigilance over every fragment could mean the difference between informed assistance and unwitting compromise.