Researchers at Manifold Security have disclosed two security weaknesses in Anthropic's Claude for Chrome extension that could allow another browser extension with access to the Claude website to trigger predefined AI-powered actions involving a user's Gmail, Google Docs and Google Calendar.
According to the researchers, the issues remain present in version 1.0.80 of the extension despite earlier mitigations introduced after the disclosure of the "ClaudeBleed" vulnerability. While Anthropic restricted how external webpages can communicate with the extension, Manifold says the underlying trust boundary that determines whether a user intentionally initiated an action has not been fully addressed.
The findings do not indicate that arbitrary websites can directly read a user's email or documents. Instead, the attack requires another browser extension that already has permission to execute scripts on the claude.ai domain. If such an extension is malicious or becomes compromised, it could abuse Claude's existing capabilities to initiate AI tasks that access a user's connected Google services.
Forged clicks can initiate predefined Claude actions
Following the earlier ClaudeBleed disclosure, Anthropic replaced unrestricted prompt handling with a fixed allowlist of predefined onboarding tasks. Rather than allowing external callers to submit arbitrary prompts, the extension now recognizes only nine task identifiers embedded within its code.
Among these are demonstration workflows for third-party services such as DoorDash, Salesforce and Zillow, along with tasks that interact with Gmail, Google Docs and Google Calendar. This design significantly narrows the attack surface because outside scripts can no longer provide custom instructions for Claude to execute.
However, Manifold Security found that the mechanism responsible for launching these tasks can still be manipulated.
The researchers explain that a content script running within the extension monitors the Claude webpage for clicks on a specific onboarding element. When a click occurs, the script reads the associated task identifier and forwards it to the extension, which opens Claude's side panel with the corresponding workflow prepared.
The problem lies in how those clicks are validated. Instead of confirming that the event originated from an actual user interaction, the extension accepts any matching click event, including one generated programmatically by JavaScript.
Modern browsers provide an "event.isTrusted" property that distinguishes genuine user actions from synthetic events created by scripts. According to Manifold, the extension does not verify this property before processing the request.
As a result, another extension capable of interacting with the Claude webpage can dynamically create the required element, assign one of the approved task identifiers and dispatch an artificial click event. Because the extension treats the event as legitimate, Claude opens the selected workflow as though the user had manually initiated it.
The researchers demonstrated this behavior using a short proof-of-concept script executed within the Claude page, showing that synthetic click events marked as untrusted were still accepted by the extension.
Approval settings determine the level of risk
Whether the forged action progresses beyond this point depends largely on how the extension has been configured.
For users operating under Claude's default "Ask before acting" setting, the extension still presents an approval prompt before carrying out actions involving Gmail, Google Docs or Google Calendar. This additional confirmation prevents automatic execution, although users could still unknowingly approve an attacker-triggered request.
The risk increases considerably for users who have enabled the optional "Act without asking" mode. In this configuration, the extension can perform supported tasks without requesting further confirmation, allowing attacker-triggered workflows to execute automatically.
Manifold assigned a CVSS severity score of 7.7 under the default approval model and 9.6 when unattended execution is enabled.
The researchers say a straightforward mitigation would be to reject any click event that was not generated by a genuine user, preventing scripts from activating these workflows through synthetic browser events.
Researchers identify second permission-handling concern
Manifold also disclosed a separate issue involving how the extension initializes permission settings when its side panel loads.
According to the researchers, if the panel starts with a specific URL parameter indicating that permission checks should be skipped, the extension immediately enters a mode that bypasses user approval for supported actions.
Although users receive a warning indicating that Claude now has broader authority to perform actions on their behalf, the privileged session has already been established by the time the notification appears.
The researchers emphasize that this second issue is not directly exploitable under current conditions because the parameter can presently be generated only by the extension itself. Nevertheless, they argue that any future vulnerability allowing a lower-privileged component to influence this parameter could eliminate the remaining approval barrier and enable silent execution.
Potential attack paths discussed by the researchers include future message-handling flaws, panel initialization bugs or cross-site scripting vulnerabilities that could expose the parameter to untrusted input.
To reduce that risk, Manifold recommends that the extension ignore permission-related values supplied through URLs and instead always initialize new sessions in approval mode.
The researchers classify the forged-task technique as an example of indirect prompt injection within the OWASP Top 10 for Large Language Model Applications because an attacker manipulates the AI agent into executing one of its own predefined workflows rather than supplying new instructions directly.
They also associate the unattended execution scenario with excessive agency, referring to AI systems that are granted broad authority to perform sensitive actions with minimal user oversight.
According to the report, these behaviors occur regardless of whether users are running Claude Opus, Sonnet or Fable, indicating that the weaknesses originate in the browser extension rather than the underlying language models.
Issues remain unresolved months after disclosure
Manifold Security reported both vulnerabilities to Anthropic on May 21 while testing version 1.0.72 of the extension. Anthropic acknowledged the reports the following day.
The forged-click issue was closed on the basis that it fell within the scope of the previously reported ClaudeBleed investigation, which Anthropic indicated remained open while a more comprehensive solution was being developed.
The permission-handling report was classified as informational because the relevant parameter was intended for workflows that users had already configured for unattended execution.
Despite those responses, Manifold says it found the same vulnerable code paths unchanged after examining version 1.0.80 released on July 7.
As of July 14, the researchers noted that no CVE identifier had been assigned to either issue and Anthropic had not published a public advisory addressing the findings.
The latest research follows a series of security concerns involving AI-powered browser agents.
Earlier this year, researchers disclosed ClaudeBleed, a vulnerability that allowed websites to inject prompts into Claude for Chrome by exploiting how the extension trusted requests originating from the Claude website itself rather than verifying which script generated them.
LayerX, which originally disclosed ClaudeBleed, described the issue as a classic "confused deputy" problem, where software possessing legitimate privileges unknowingly performs actions on behalf of an untrusted requester.
Security researchers have also identified comparable trust-boundary weaknesses affecting other Anthropic products, including Claude Code, demonstrating broader challenges associated with AI agents that can directly interact with browsers, developer environments and online accounts.
The latest findings reinforce the importance of carefully validating user intent before granting AI assistants access to sensitive online services. As AI-powered browser agents become increasingly capable of interacting with email, documents and productivity platforms, researchers argue that ensuring those actions genuinely originate from users remains one of the most critical security controls.
.jpg)
