Threat Intelligence
GTIG Report: Adversarial AI Shifts to Agentic Workflows and Automated Credential Harvesting
Google Threat Intelligence Group reports that threat actors are moving from basic prompting to fully agentic AI workflows, significantly compressing defender response windows. Recent incidents include automated mass credential harvesting executed in under six hours, alongside rising attacks targeting proprietary AI intellectual property and software supply chains.
What Happened
Google Threat Intelligence Group (GTIG) has released its Q2 2026 findings on adversarial AI, highlighting a major operational shift from basic prompt-based misuse to fully agentic AI workflows and automated execution. This transition drastically reduces human-in-the-loop latency, shrinking the operational window defenders rely on to detect and remediate threats.
Notably, GTIG observed an adversary compromise a cloud resource and subsequently leverage agentic automation to plan, build, and execute a mass credential harvesting campaign in less than six hours. Concurrently, groups like UNC6780 have begun actively manipulating AI coding assistants and large language model (LLM) security scanners to slip malicious code into open-source software supply chains.
Impact
Organizations deploying enterprise AI solutions face expanded threat vectors targeting both the development lifecycle and runtime infrastructure. Key impact areas identified in the report include:
- Accelerated Attack Chains: Autonomous multi-step operations reduce detection opportunities for traditional SecOps monitoring.
- Supply Chain Poisoning: Manipulation of AI-assisted coding tools and LLM security scanners bypasses standard automated code reviews.
- IP and Compute Theft: Attackers are directly targeting proprietary models, source code, and API credentials across healthcare, government, and media sectors while co-opting victim cloud resources to run unauthorized AI workloads.
What to Do
Defenders and system administrators must adapt their monitoring and governance strategies to address autonomous and AI-centric threats:
- Audit and restrict cloud compute quotas to quickly detect unauthorized, high-intensity AI workloads.
- Implement strict access controls and monitoring for proprietary model weights, internal source repositories, and API credentials.
- Validate the output of AI coding assistants and LLM security scanners using traditional, non-AI code review gates and continuous software composition analysis.
- Harden developer workstations and internal environments against threat actors attempting to exploit automated pipelines.
Related articles
Astrana Health Discloses Data Breach Stemming from Employee Social Engineering
Astrana Health reported a data breach impacting private and confidential records after attackers used targeted social engineering, including phone number spoofing, to impersonate internal personnel and access company servers. The healthcare management firm has enacted credential rotations, restricted remote access tools, and initiated system restorations from clean backups.
Read briefing
Unit 42 Analysis: State of AI-Enabled Malware and Endpoint Detection Efficacy
Recent threat intelligence research into 405 AI-enabled malware samples reveals that roughly 97% remain confined to public repositories and sandboxes without reaching production environments. Standard behavioral analysis, endpoint detection telemetry, and cloud-based sandboxing successfully intercept and block these threats using existing rule sets, proving that AI integration alters authoring methods rather than fundamental execution mechanics.
Read briefing
Continuous Authorization Challenges: How SecOps Can Bridge the Zero Trust Gap
Organizations struggle to enforce continuous authorization while maintaining user productivity. The lack of automated, context‑aware controls leads to blind spots in identity and access management. Defenders should adopt AI‑driven, agentic SOC workflows to enforce real‑time policy checks and reduce risk.
Read briefing
External Threat Intelligence Trends: LLM-Generated Malware and SOC Automation Analysis
Industry analysis for Q3 2026 highlights the rise of LLM-generated threats like PhantomRaven and the shift toward automated Agentic SOC models. Defenders are advised to integrate external intelligence directly into triage workflows to counter accelerating exploitation windows.
Read briefing