Skip to navigation

Threat reportVulnerabilityTL-2026-1380

"PromptFiction" Claude Desktop Auto-Submit Flaw Chained With "Claudy Day" Claude.ai Exploit Chain Enables Silent Exfiltration and, via Filesystem Server MCP, Local RCE

highPATCHED

Oasis Security disclosed PromptFiction, a flaw in Anthropic's Claude Desktop `claude://` URI-scheme handler that auto-submitted attacker-crafted prompts to Claude with a single click and zero user review.

Key facts

IdentifiersTL-2026-1380
Also known asPromptFiction, Claudy Day
SeverityHIGH
Affected productsAnthropic Claude Desktop; Anthropic Claude.ai (web); Anthropic Filesystem Server (official MCP integration)
Weaknesses (CWE)CWE-77, CWE-79, CWE-601, CWE-306, CWE-200
ToolsAnthropic Filesystem Server (official MCP integration)
MITRE ATT&CK techniquesT1027 Obfuscated Files or Information, T1059 Command and Scripting Interpreter, T1071.001 Web Protocols, T1074.001 Local Data Staging, T1083 File and Directory Discovery, T1119 Automated Collection, T1140 Deobfuscate/Decode Files or Information, T1189 Drive-by Compromise, T1204.001 Malicious Link, T1213 Data from Information Repositories, T1505 Server Software Component, T1528 Steal Application Access Token and 10 more in the attack flow
MotivationUNKNOWN
Target sectorstechnology, software-development, any-sector-using-ai-agents
Target regionsGlobal
Detection rules9 (1 shown free)
Sources6 cited sources

Detection coverage

As of 2026-07-15, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-1380 across Splunk SPL, Microsoft KQL and Sigma, covering 16 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: claude:// URI Auto-Submit Launch From Malvertised/Redirected Link (Exploitation)

Detects Claude Desktop being launched via the claude:// custom URI scheme from a browser process shortly after navigation through the claude.com/redirect open-redirect endpoint or an ad-network/email referrer, matching the PromptFiction one-click auto-submit exploitation stage.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1204.001, T1566.002, T1583.008
  • Data sources: sysmon, windows_security, proxy_logs
  • Severity: high
  • Confidence: high

Splunk SPL

index=endpoint sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" EventCode=1
| eval cmdline=lower(CommandLine)
| where like(cmdline, "%claude://%")
| eval parent=lower(ParentImage)
| where like(parent, "%chrome.exe%") OR like(parent, "%msedge.exe%") OR like(parent, "%firefox.exe%") OR like(parent, "%outlook.exe%")
| join type=left ParentProcessId [ search index=proxy sourcetype="proxy:web" (uri_path="*claude.com/redirect*" OR uri_path="*claude.ai/new*q=*") earliest=-5m
| stats count as redirect_hits by ParentProcessId ]
| table _time, ComputerName, User, cmdline, parent, redirect_hits
| sort -_time

Microsoft KQL

DeviceProcessEvents
| where Timestamp > ago(24h)
| where ProcessCommandLine has "claude://"
| where InitiatingProcessFileName in~ ("chrome.exe","msedge.exe","firefox.exe","outlook.exe")
| project Timestamp, DeviceName, AccountName, ProcessCommandLine, InitiatingProcessFileName, InitiatingProcessCommandLine
| join kind=leftouter (
    DeviceNetworkEvents
    | where Timestamp > ago(24h)
    | where RemoteUrl has_any ("claude.com/redirect", "claude.ai/new")
) on DeviceName
| project Timestamp, DeviceName, AccountName, ProcessCommandLine, InitiatingProcessFileName, RemoteUrl

Sigma

title: claude:// URI Handler Invocation Following Redirect or Ad-Network Referrer
id: 7c1e4a2e-2f4a-4c2f-9a1e-5b1c3d0a1001
status: experimental
description: Detects Claude Desktop's claude:// URI scheme being invoked from a browser or mail client process, consistent with the PromptFiction one-click auto-submit exploitation stage delivered via malvertising, Gmail ads, or SEO-poisoned search results.
references:
    - https://www.oasis.security/blog/claude-ai-prompt-injection-data-exfiltration-vulnerability
    - https://hackread.com/promptfiction-flaw-auto-prompts-claude-desktop/
author: Threadlinqs Team
date: 2026-07-15
tags:
    - attack.initial-access
    - attack.t1204.001
    - attack.t1566.002
    - attack.t1583.008
    - attack.execution
logsource:
    category: process_creation
    product: windows
detection:
    selection_uri:
        CommandLine|contains: 'claude://'
    selection_parent:
        ParentImage|contains:
            - 'chrome.exe'
            - 'msedge.exe'
            - 'firefox.exe'
            - 'outlook.exe'
    condition: selection_uri and selection_parent
falsepositives:
    - Legitimate developer clicking a trusted internal claude:// deep link
level: medium

Known false positives:

  • Internal/trusted claude:// deep links used by developers for legitimate automation

Tuning notes: Scope the parent-process allowlist to your organization's approved browsers/mail clients and add a lookup of known-benign internal claude:// launchers to reduce noise; tighten the redirect join window if proxy logging lag is high.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Hidden Below-the-Fold Prompt Payload Delivered via claude.ai/new?q= Parameter (Exploitation) · unlock with TL Creds
  • Remote-Debug Code Insertion Into Project Source Files by MCP Filesystem-Server Process (Post-Exploitation) · unlock with TL Creds
  • Conversation-History/Memory Search Followed by Local File Staging and Upload Attempt (C2/Exfiltration) · unlock with TL Creds
  • claude.com Open-Redirect Used to Cloak claude:// or claude.ai/new Destination (Exploitation) · unlock with TL Creds
  • Filesystem-Server MCP Process Writing Outside Its Approved Directory Scope (Post-Exploitation) · unlock with TL Creds
  • Sandbox Egress to [indicator] Using a Non-Session API Key (C2/Credential Abuse) · unlock with TL Creds
  • PowerShell/Shell Profile Write by AI-Agent or MCP Filesystem-Server Process (Post-Exploitation Persistence) · unlock with TL Creds
  • Outbound Files-API Upload to [indicator] Immediately Following Untrusted URL Navigation (C2/Exfiltration) · unlock with TL Creds

Indicators of compromise

As of 2026-07-15, Threadlinqs Intelligence tracks 16 indicators of compromise for TL-2026-1380.

Sample indicators (1 of 16)

  • Domain: api[.]anthropic[.]com

Defanged, TLP:CLEAR. Full set of 16 indicators: unlock with TL Creds.

16 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

Chained with the previously disclosed three-vulnerability "Claudy Day" attack path (invisible URL-based prompt injection, Files-API data exfiltration, and a claude.com open redirect), the combined chain enabled silent exfiltration of prior conversation history and, when Anthropic's official Filesystem Server MCP integration is installed, local file read/write, persistence, and remote code execution on the victim machine. Anthropic fixed both issues through its Responsible Disclosure Program; the desktop fix shipped in Claude Desktop v1.1.2321.

Oasis Security's research lead Elad Luz documented a two-stage exploit chain against Anthropic's Claude AI product family. The first stage, "Claudy Day" (published 2026-03-18, updated 2026-05-27), consisted of three chained web-based flaws in Claude.ai: (1) invisible prompt injection via the `claude.ai/new?q=` URL parameter, where certain HTML tags embedded in the query string were invisible in the rendered text box but were fully parsed and processed by Claude once the user pressed Enter; (2) data exfiltration through Anthropic's own Files API, abused because Claude's code-execution sandbox blocks arbitrary outbound network access but explicitly allow-lists `api.anthropic.com` — an attacker could smuggle an API key for their own Anthropic account inside the hidden prompt and instruct Claude to search the user's conversation history/memory, write the sensitive findings to a file inside the sandbox, and upload that file to the attacker's account via the Files API; and (3) an open redirect at `claude.com/redirect/<target>` that let attackers disguise the malicious `claude.ai/new?q=` link as a trusted claude.com URL, which was in turn used to get the link approved by ad-network URL validation (which checks hostname only) so it could be purchased as a Google Ads / Gmail ad result targeting selected victims. Oasis stressed that the Claudy Day chain worked against a completely default, bare-bones Claude session — zero integrations, zero tools, zero MCP servers required.

The second stage, PromptFiction (published 2026-07-15), targeted the Claude Desktop application specifically. Claude Desktop registers a custom OS-level URI scheme, `claude://`. Before the fix, a crafted link such as `claude://claude.ai/new?q=<prompt>` would open Claude Desktop and immediately submit the embedded prompt to the agent with a single click and no confirmation screen — unlike the web client, which displays the pre-filled prompt in the compose box pending explicit user send. This let an attacker skip the "user reviews then sends" gate entirely on the desktop app. To reduce operator suspicion, attackers could front-load a benign, friendly request (e.g., "tell me a joke") at the start of the visible prompt text and push the actual malicious instructions further down, exploiting Claude's UI behavior of collapsing long messages behind a "show more" toggle — the harmful payload sat below the fold and was processed in full even though the victim never expanded it. Oasis noted the malicious `claude://` link could be concealed using the same claude.com open-redirect technique documented in Claudy Day, so the final delivery vector again looked like a trusted claude.com address while actually launching the desktop app with attacker-controlled instructions, and could again be distributed via Google Search/Gmail ads aimed at chosen targets.

Chaining the two: a single click from a malicious ad or link opened Claude Desktop and auto-submitted a hidden prompt (PromptFiction) that carried out the Claudy Day exfiltration logic (conversation-history theft via the Files API) with no further interaction. Where the victim's Claude Desktop had Anthropic's official Filesystem Server MCP integration enabled (used to grant Claude read/write access to approved local directories for coding and file-management workflows), the injected prompt could instead direct Claude to write or modify files within those approved directories — for example inserting remote-debugging code into Python or JavaScript project files during what looked like a routine coding task, or appending persistence hooks into shell rc files (`.zshrc`, `.bashrc`) reachable under the approved filesystem scope — escalating the impact from conversation-data theft to local file read/write, persistence across reboots/shell sessions, and ultimately remote code execution on the victim's machine, all without any confirmation dialog beyond the single initial click.

No CVE identifier or CVSS score was assigned to either the Claudy Day or PromptFiction disclosures at time of publication. Anthropic remediated both issues through its Responsible Disclosure Program: the prompt-injection/auto-submit behavior in Claude Desktop was fixed in version 1.1.2321, which now requires the user to manually review and explicitly send any prompt delivered via the `claude://` URI scheme rather than auto-submitting it; the Claude.ai web-side prompt-injection vector from Claudy Day was likewise patched, with Oasis noting some related hardening work continued after initial disclosure.

MITRE ATT&CK techniques by tactic:

Defense Evasion

T1027 Obfuscated Files or Information; T1140 Deobfuscate/Decode Files or Information; T1564 Hide Artifacts

Execution

T1059 Command and Scripting Interpreter; T1204.001 Malicious Link

Command and Control

T1071.001 Web Protocols

Collection

T1074.001 Local Data Staging; T1119 Automated Collection; T1213 Data from Information Repositories

Discovery

T1083 File and Directory Discovery

Initial Access

T1189 Drive-by Compromise; T1566.002 Spearphishing Link

Persistence

T1505 Server Software Component; T1546.004 Unix Shell Configuration Modification

Credential Access

T1528 Steal Application Access Token

Impact

T1565.001 Stored Data Manipulation

Exfiltration

T1567 Exfiltration Over Web Service; T1567.002 Exfiltration to Cloud Storage

Resource Development

T1583.008 Malvertising; T1585.003 Cloud Accounts

Reconnaissance

T1589 Gather Victim Identity Information

stealth

T1684.001 Impersonation

Remediation

Patches

  • Claude Desktop v1.1.2321 — claude:// URI-scheme prompts now require manual user review/send instead of auto-submitting

Immediate actions

  • Update Claude Desktop to version 1.1.2321 or later, where claude:// links are pre-filled but require explicit manual review and send rather than auto-submitting
  • Treat any claude:// or claude.ai/new?q= link received via email, ads, or search results as untrusted until the prompt content has been fully reviewed before sending
  • Audit which local directories are exposed to Claude Desktop via the official Filesystem Server MCP integration and scope them to the minimum necessary paths
  • Review shell rc files (.zshrc, .bashrc) and project source files in any Claude-Filesystem-Server-approved directory for unexpected modifications, persistence hooks, or inserted remote-debugging code

Workarounds

  • Disable or do not install the Filesystem Server MCP integration on machines used for sensitive work until Claude Desktop is fully patched
  • Do not click claude:// or claude.ai links delivered via advertisements, unsolicited email, or unfamiliar search results

Longer-term hardening

  • Require explicit user approval before any AI-agent action that reads conversation history/memory and sends it to an external API, even first-party APIs
  • Sanitize and validate all URL-based prompt inputs server- and client-side, including detection of hidden/invisible HTML or encoded content embedded in query parameters
  • Audit and restrict code-sandbox outbound network allow-lists so a first-party API domain cannot be repurposed as an attacker-controlled exfiltration channel via attacker-supplied credentials
  • Fix open-redirect endpoints (e.g., claude.com/redirect/<target>) to validate destination hostnames/paths against an allow-list rather than redirecting unconditionally
  • Monitor MCP integrations (Filesystem Server and similar) for anomalous file writes, shell-config modifications, or code insertions that follow an AI agent session

Affected and fixed versions

  • Anthropic — Claude Desktop
    Vulnerable versions: < 1.1.2321
    Fixed in: 1.1.2321
  • Anthropic — Claude.ai (web)
    Vulnerable versions: pre-patch, prior to 2026-05-27 hardening
    Fixed in: patched via Responsible Disclosure Program, prompt-injection vector closed
  • Anthropic — Filesystem Server (official MCP integration)
    Vulnerable versions: any version when installed alongside vulnerable Claude Desktop
    Fixed in: mitigated once Claude Desktop 1.1.2321 closes the auto-submit vector

Timeline

  • Oasis Security releases a companion full technical report PDF ('locked-ioc-2') alongside the Claudy Day blog post, describing malicious-search-result placement as an additional delivery vector for the injected claude.ai/new?q= link, alongside malvertising.
  • Oasis Security publishes the Claudy Day research, documenting a three-vulnerability chain in Claude.ai: invisible URL-based prompt injection, Files-API data exfiltration, and a claude.com open redirect, working against a fully default bare-bones Claude session.
  • Oasis Security updates the Claudy Day blog post noting the prompt-injection vector has been fixed by Anthropic while other related hardening work continues.
  • Oasis Security discloses in the PromptFiction writeup that the finding was a duplicate: an independent researcher had separately reported the same Claude Desktop claude:// auto-submit issue to Anthropic beforehand but chose not to publish their own findings.
  • Hackread publishes a technical breakdown of the PromptFiction claude:// auto-submit mechanism and its concealment techniques (front-loaded benign text, 'show more' UI collapse abuse).
  • Dark Reading reports on the PromptFiction/Claudy Day chained vulnerability disclosure, summarizing the exfiltration and RCE-via-MCP risk for enterprise readers.
  • Anthropic ships Claude Desktop version 1.1.2321, requiring users to manually review and send any prompt delivered via the claude:// URI scheme rather than auto-submitting it, closing the PromptFiction auto-submit vector.
  • Oasis documents that PromptFiction can be chained with Claudy Day's open-redirect and exfiltration techniques, escalating impact to silent conversation-history theft and, with Filesystem Server MCP installed, local file read/write, persistence, and remote code execution.
  • Oasis Security research lead Elad Luz publishes the PromptFiction finding: a click-to-auto-submit flaw in Claude Desktop's claude:// URI-scheme handler that bypasses the prompt-review step present on the web client.

Sources cited

Previous and next in vulnerability

Drafted by the Threadlinqs research pipeline from 6 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats