Cyble says agentic AI can investigate phishing, with human sign-off

What happened

According to Cyble, in a blog post dated November 12, 2026, an agentic AI can run most of a phishing investigation on its own, but a human should sign off before anything is acted on. Cyble says the agent can pull domain registration records, open the lure safely, pivot to related infrastructure, extract indicators and write a verdict with evidence attached. Cyble

This is a vendor blog post, not a vulnerability disclosure, and there is no CVE or patch. The claims below are Cyble’s, and we have not independently verified them.

• Cyble says Barracuda reported on Oct. 7 that one phishing email can carry a lure for the human and a hidden prompt for the AI reading it. The Barracuda report itself was not among our sources.
• Cyble says Microsoft’s randomized trial of 167 security analysts, on the Security Copilot Phishing Triage Agent, found analysts were faster and more accurate. It reports up to 6.5 times as many true positives per minute and up to 77% higher accuracy by F1 score. Cyble adds that Microsoft studied its own product and the test measured triage of single emails, not whole campaigns.
• Cyble says Gartner predicted in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027, and estimated only about 130 vendors offer real agentic capabilities.
• Cyble walks through a reconstructed investigation of lookalike Microsoft Teams domains that researcher Steven Lim posted in October 2026. It says the seed domain was teams-online[.]com, registered Oct. 4, 2026, per Lim’s screenshot. Cyble states the walkthrough is built from Lim’s public artifacts and is not output from a specific product run.

Not confirmed: the later steps of the walkthrough and Cyble’s full list of questions to ask before trusting an agent were cut off in the source text we received. Whether any victim was affected is also not confirmed.

Am I affected?

Only if your team or tooling uses an AI agent that reads untrusted content, such as reported phishing emails, web pages or domain data, and then chooses its own next steps or acts on its verdict. Cyble says that autonomy is what makes agents useful and what makes them open to being steered by content they read. Cyble

If you do not use such agents, the sources describe no action for you. No affected product versions are named, and exploitation in the wild is not confirmed here.

What to do now

1. Check whether any AI agent in your security workflow reads attacker-controlled content, such as phishing emails or lure pages.
2. Follow Cyble’s stated position: do not let the agent act on its verdict unchecked. Require a human to review the evidence trail before any blocking, takedown or user notification.
3. Treat the vendor figures above as claims. Cyble notes the Microsoft trial covered the vendor’s own product and single-email triage, not multi-day campaign investigations.
4. If you evaluate an agent product, test it on a real case and check that it shows the evidence behind each verdict. Cyble’s full evaluation checklist is not confirmed in the text available to us, so read the original post.

Sources

• Cyble: Can Agentic AI Investigate a Phishing Campaign? Evidence Walkthrough