Zscaler found hidden web-page instructions that tricked four of 26 AI models into paying an attacker's crypto wallet
At a Glance
- Significant disruption. Hours of lost work, reputational embarrassment, or measurable financial impact under $10k.
Independent project · aggregated from public reports and may be unverified — see the primary source below · not affiliated with or endorsed by any company or product named.
What Happened
Zscaler ThreatLabz documented two live campaigns using indirect prompt injection to manipulate AI agents browsing the web. The first impersonated a Python library and hid instructions in page content directing the agent to buy a non-existent developer API license priced at $3.00. Tested across 26 large language models, four failed to take appropriate action and executed the fraudulent payment, transferring roughly 0.0012 ETH to an attacker-controlled wallet. A second campaign targeting users seeking a cryptocurrency portfolio tracker caused two models to classify the fraudulent site as legitimate in some contexts. Human visitors to the same sites were shown equivalent card and crypto payment scams.
Case Analysis
Verified Facts
- Four of 26 tested models executed a fraudulent cryptocurrency payment after indirect prompt injection
- The fake product was a non-existent developer API license priced at $3.00, paid as about 0.0012 ETH
- A second campaign caused two models to classify a fraudulent site as legitimate
Not Publicly Confirmed
- How much money real users lost to these campaigns
- Which specific models failed
Operational Lessons
- Agents with payment authority need an out-of-band human confirmation for every transfer
- Web content an agent reads is attacker-controlled input, not reference material
Primary Source
Indirect Prompt Injection in Web Content Targets AI Agents (Zscaler ThreatLabz)zscaler.com ↗Case Record
More Cases
Prompt injection reached host-level code execution in Microsoft Semantic Kernel through eval() and a stray annotation
Microsoft disclosed two vulnerabilities that turn prompt injection into host compromise in its Semantic Kernel agent framework, which has over 27,000 GitHub stars. CVE-2026-26030 affects the Python package before 1.39.4: the default in-memory vector store filter is a Python lambda executed with eval() on unsanitized model-controlled input, so an attacker could escape the template string, traverse Python's class hierarchy, bypass the AST blocklist validator and run arbitrary commands, demonstrated by launching calc.exe from a single prompt injection. CVE-2026-25592 affects the .NET SDK before 1.71.0: DownloadFileAsync was accidentally marked with a [KernelFunction] attribute, exposing it to the model with an entirely AI-controlled, unvalidated local file path, allowing writes to locations such as the Windows Startup folder and thus sandbox escape.
SearchLeak let a single click on a malicious link pull emails, files and MFA codes out of Microsoft 365 Copilot
Varonis Threat Labs disclosed CVE-2026-42824, a critical flaw in Microsoft 365 Copilot Enterprise Search. Clicking a single crafted link chained three bugs: injection of the URL parameter into the prompt, a rendering race condition, and abuse of Content Security Policy allowlisting to exfiltrate data through Bing's infrastructure. The attack could reach emails and calendar details, indexed SharePoint and OneDrive files, one-time and MFA codes, password-reset links, meeting notes and salary data. Because Copilot Enterprise is a managed service, tenant admins could not patch it themselves; Microsoft mitigated the flaw on its backend.
Cursor's command allowlist could be bypassed with shell built-ins, giving prompt injection a silent path to code execution
Pillar Security disclosed CVE-2026-22708 in Cursor. In Auto-Run Mode with an allowlist enabled, shell built-ins such as export, typeset, declare, readonly, unset and local were implicitly trusted by Cursor's server-side evaluator and executed without appearing in the allowlist or requiring approval, because they run inside the shell session rather than as separate binaries. An attacker delivering indirect prompt injection could silently poison environment variables and then trigger malicious code through trusted developer tools, producing both zero-click and one-click remote code execution. Pillar reported it in August 2025, Cursor acknowledged it as a systemic issue in September 2025, and the fix shipped in version 2.3 in January 2026, which now requires explicit approval for any command the parser cannot classify.