TAU-HOME.COM
LOADING

21 Zero-Click Prompt Injection Attacks Without CVEs: Limits of Egress Whitelists and Production AI Security Threats

An analysis of at least 21 publicly documented prompt-injection attacks and key vulnerabilities (such as LiteLLM and Gortex) that hit production AI assistants f

tau · November 21, 2026

#Prompt Injection #AI Security #Vulnerability #MCP #CVE

21 Zero-Click Prompt Injection Attacks Without CVEs: Limits of Egress Whitelists and Production AI Security Threats

Concept diagram of zero-click prompt injection attack structures and data exfiltration paths targeting production AI assistants

Image source: https://konsole.one/threat-intel/prompt-injection-in-the-wild-2025-2026

Between 2025 and November 2026, security threats targeting commercial artificial intelligence (AI) assistants—conversational tools adopted by enterprises and individual users worldwide—surged significantly. At least 21 publicly documented prompt-injection attacks (security vulnerabilities that manipulate AI models into unintended behaviors through hidden text instructions) were uncovered across production environments from Microsoft, OpenAI, Google, and Salesforce. Notably, only three of these incidents received official Common Vulnerabilities and Exposures (CVE) identifiers, underscoring how traditional security frameworks continue to miss AI-specific risks.

Concept diagram of zero-click prompt injection attack structures and data exfiltration paths targeting production AI assistants

Image Credit: Konsole Threat Intel

These security findings reveal the structural vulnerabilities lurking behind the rise of AI agents (AI systems that autonomously handle tools and files based on user prompts). In particular, because the attacks evolved into "zero-click" vectors—where malicious payloads hide inside ordinary user queries to exfiltrate private data without requiring any manual execution steps from users or administrators—heightened vigilance is required.

The Common Thread of 21 Attacks: The Ambiguity of Data Versus Instructions

The 21 prominent production attacks tracked by security researchers largely followed a consistent script. Attackers planted malicious text wherever an AI assistant might later read it: emails, calendar invites, shared documents, public web pages, GitHub issues, and web-to-lead form fields. The text was disguised using techniques such as white-on-white typography or hidden HTML comments, making it invisible to human readers.

Days later, when a victim asked the assistant an unrelated, ordinary question, the problem manifested. During retrieval, the hidden malicious text was pulled into the context window (the working memory space processed by the model at once). Because current AI architectures struggle to strictly separate data from instructions, the model treated the text as legitimate user instructions and executed it.

  • Private Data Exfiltration: The AI located the most sensitive information within reach (API keys, personal notes, internal documents) and embedded them into URLs controlled by the attacker.
  • Diverse Exfiltration Channels: Every available outbound communication vector—markdown image loading, browser tool calls, Domain Name System (DNS) lookups, and outbound emails—was exploited to carry data out.

Structural Limits of Egress Whitelists and Key Breaches

Most organizations rely on egress allowlists to control outbound network traffic from AI servers. However, this defense mechanism was bypassed in at least five major incidents.

  • In the Microsoft 365 Copilot "EchoLeak" incident (CVE-2025-32711), Microsoft's own Teams and SharePoint proxy domains were trusted by the system, allowing data to escape through proxy URLs.
  • In the Salesforce Agentforce "ForcedLeak" incident, an attacker purchased an expired domain for five dollars that still sat on Salesforce's trusted URL list, successfully intercepting customer relationship management (CRM) data.
  • Zero-click attacks via calendar invites and Google Drive documents were also confirmed in ChatGPT Connectors and Google Gemini. For Gemini, researchers at Black Hat demonstrated scenarios where compromised commands manipulated smart home devices to turn off lights or activate water boilers.

The 2026 Turning Point: Coding Agents and Software Supply Chain Compromises

While 2025 marked the year of data exfiltration targeting enterprise copilots, February 2026 saw attacks escalate into actual software supply chain compromises.

Cline, a popular open-source coding agent, featured a GitHub issue-triage workflow that fed issue titles verbatim into model prompts. An attacker exploited this by crafting a malicious issue title that hijacked the workflow, poisoned the GitHub Actions cache, and ultimately stole the Node Package Manager (npm) publishing token. On February 17, 2026, a compromised version of the package was published for roughly eight hours, installing unwanted agents on developer machines during routine updates.

Security firms like Snyk and the Cloud Security Alliance described this as the first documented case of indirect prompt injection against an AI agent being successfully leveraged to compromise a software supply chain.

Implications and Limitations for AI Security

These incidents demonstrate that traditional network-perimeter security defenses alone cannot adequately mitigate the threats posed by autonomous AI agents. As OpenAI's Chief Information Security Officer noted, prompt injection remains "a frontier, unsolved security problem."

Unless foundational architectural improvements allow AI systems to reliably distinguish untrusted data from verified user instructions, organizations and developers face the ongoing challenge of guarding against zero-click risks across every agent pipeline handling external inputs.

Sources

Impact-Site-Verification: 9b3ebd10-6c45-44cf-a014-5b8c19b7b285