Feb 13, 2026 — Google's security team warned in a technical report: multiple hacking groups are abusing the Gemini AI for reconnaissance and attack support. The report details show attackers are not merely treating the AI as a writing aid but are using Gemini as a programmable "recon engine" and "attack assistant."
Google noted several concrete techniques in the attack chain: first, attackers use structured queries over public information and victims' historical communications to rapidly generate highly tailored personal profiles, including job descriptions, common signature formats, frequent collaborators, and typical repository paths. Based on these outputs, attackers automate the construction of phishing emails and forged login pages, producing multiple copy variants for A/B testing and significantly improving success rates.
Second, the report describes how phishing kits are generated and iterated: attackers leverage Gemini to quickly produce webpage templates, social‑engineering copy, and instructions prompting victims to upload credentials, then assemble these components onto ready‑made hosting and short‑lived domains for distribution. Google observed some samples containing AI‑generated images and customized logos to increase deceptive credibility.
The third category of abuse involves malware "staging" and obfuscation techniques. Attackers have used the model to write and revise payloads, generate obfuscated code snippets to evade static detection, and even employ Gemini to parse defensive logs and optimize backdoor persistence procedures. The report also mentions "model extraction" attempts, where adversaries try to reconstruct behavior and parameters of internal or restricted models through extensive queries to replicate their capabilities.
Google included actionable detection indicators and defensive recommendations: strengthen API and model access controls, create alerts for anomalous queries and high‑frequency interactions, require human review for queries containing sensitive context, rotate credentials regularly, and limit automated access. Google also provided example rules and IOCs to help administrators deploy protections in GCP and Workspace environments.
The Hacker News republished and supplemented the report the same day, noting researchers observed consistent abuse patterns across multiple samples, although the public report does not fully attribute specific groups to particular countries. Google emphasized that the stronger a model's capabilities, the greater the risk of misuse, and operators and enterprises must balance productivity gains with technical and policy controls to limit abuse.
The report presents a practical challenge for security teams: when AI can generate personalized attack content and automatically test variants in minutes, traditional signature‑based defenses become more brittle. Google recommends focusing on behavior‑based detection, access governance, fast incident response processes, and more targeted anti‑phishing training for employees.
For technical staff and managers concerned with AI security and enterprise protection, the Feb 13, 2026 report is both a warning and a blueprint: as AI drives productivity gains, related abuse scenarios have moved from concept to practice and must be met with finer‑grained monitoring and policy measures.
▶️ Enjoy this article? Creating takes real effort!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions, and shares are our greatest support! 🙏
More AI news at: https://www.airosetta.com