Darktrace Says It Quarantined a Prompt-Injection Email Before a Gemini Agent Could Process It
Summary
In Darktrace’s test, Darktrace / EMAIL quarantined a hidden-instruction email before it reached a Gemini agent. The vendor says it used language and sender-context signals, rather than malware, links, attachments, or blacklist matches.
Key points
- Darktrace says a test email containing hidden prompt instructions was quarantined before a customer’s Gemini agent could process it.
- The test email reportedly had no malware, links, or attachments and passed traditional validation checks.
- Darktrace / EMAIL scored the message as a 93/100 anomaly, citing machine-directed language and sender context, including a freemail address at yahoo[.]com and no prior mail history with the recipient.
- The article describes the test as a vendor demonstration; it does not report independent validation of the detection result.
- Darktrace cites EchoLeak (CVE-2025-32711), a zero-click Microsoft 365 Copilot vulnerability, as a related example of email-based prompt injection risk.
- The article recommends combining model guardrails with behavioral monitoring and investigating unusual agent actions.
Article Details
- Topic
- Behavioral detection of email-based prompt injection attacks targeting enterprise AI agents
MITRE ATT&CK
CVE
Threat Actors
Vendors
AnthropicAnthropic disclosed in November 2025 that a Chinese state-sponsored group they tracked as GTG-1002 ran the first documented AI-orchestrated espionage campaign against ~30 targets by posing as employees of a legitimateDarktraceDarktrace detected and quarantined a prompt injection email before a Gemini AI agent in Google Cloud could process it, showing that natural-language attacks can bypass traditional signature-based defenses.GoogleDarktrace detected and quarantined a prompt injection email before a Gemini AI agent in Google Cloud could process it, showing that natural-language attacks can bypass traditional signature-based defenses.MicrosoftEchoLeak (CVE-2025-32711) is cited as a related zero-click Microsoft 365 Copilot vulnerability that enabled data exfiltration from email content.
Products
Claudefor delivery through email, setting up the malicious instructions before the AI agent processed them (‘Claude generated a prompt injection payload’).Darktrace / EMAILDarktrace / EMAIL quarantined a prompt injection email before it reached an enterprise Gemini AI agent.Darktrace / SECURE AIWhile Darktrace / SECURE AI is designed to prevent agents from producing unintended outcomes, we wanted to see if we could catch and prevent this threat type earlier in the attack life-cycle, at the email security layer.GeminiDarktrace detected and quarantined a prompt injection email before a Gemini AI agent in Google Cloud could process it, showing that natural-language attacks can bypass traditional signature-based defenses.Google CloudDarktrace detected and quarantined a prompt injection email before a Gemini AI agent in Google Cloud could process it, showing that natural-language attacks can bypass traditional signature-based defenses.Microsoft 365 CopilotEchoLeak (CVE-2025-32711) is cited as a related zero-click Microsoft 365 Copilot vulnerability that enabled data exfiltration from email content.