CSA paper on GREYVIBE and AI narrowing the gap between state-adjacent and elite cyber actors

In two minutes

• The paper argues that generative AI is shrinking the gap between state-adjacent groups and elite nation-state programs. It says AI raises the floor without leveling the ceiling (p. 11).
• Its case study is GREYVIBE, a Russia-nexus cluster that WithSecure Labs disclosed on May 28, 2026. It has targeted Ukrainian military, government and civilian organizations since at least August 2025 (pp. 4, 6-7).
• WithSecure rates the group low to moderately sophisticated, with repeated operational security failures. The paper says AI (ChatGPT, Gemini, Ideogram) covered enough skill gaps to sustain five campaigns and a custom malware set (pp. 7-9).
• The paper cites other sources for the wider trend: AI-enabled malware that queries live models (LAMEHUG, PROMPTFLUX), higher phishing click rates, and an Anthropic-reported campaign that AI ran largely on its own (pp. 9-12).
• It says attribution gets harder. AI-generated code may carry less authorship signal, and states buy access from criminal markets (p. 13).
• For enterprises it recommends revisiting threat models more often, including state-adjacent actors, and shifting to behavioral detection. It also recommends Zero Trust and monitoring of internal AI tool use (pp. 14-15).
• It ends with a counterweight from Mandiant's M-Trends 2026: most intrusions still come from basic failures, and AI mostly amplifies existing attack patterns (p. 18).
• The document says it was generated with AI assistance and has not gone through official CSA review (p. 2).

Key numbers

• 89%: Year-over-year increase in attacks attributed to AI-enabled adversaries, per CrowdStrike's 2026 report as cited (p. 4).
• more than 1,100: New threat groups emerging in 2025, per ISACA as cited (p. 4).
• 80-90%: Share of a Chinese state-sponsored espionage campaign that AI handled autonomously, per Anthropic as cited (pp. 4-5).
• approximately 30: Global targets in that campaign (pp. 11-12).
• 54% vs 12%: Click-through rate of AI-generated phishing versus non-AI phishing, per a Microsoft April 2026 analysis as cited (p. 10).
• 51% at $2.77 per exploit: CVE-Genie reproduction rate for 2024-2025 CVEs and average cost per exploit (p. 9).
• 78%: Share of compromised environments with over-permissioned AI agents, per Five Eyes agentic AI guidance as cited (p. 14).
• more than 60: Iran-aligned cyber groups said to mobilize within hours of a geopolitical trigger (p. 12).

What to do with it

• Revisit threat models on a shorter cycle and add state-adjacent actors with AI-compensated tradecraft, even if your organization was previously below the targeting threshold (p. 14).
• Stop relying on poor writing or cultural errors as phishing indicators. Invest in behavioral detection of anomalous authentication, data access and process chains (p. 14).
• Reduce blast radius by limiting over-permissioned AI agents and applying Zero Trust: continuous verification, micro-segmentation, behavioral analytics (p. 14).
• Update detection engineering for AI-generated artifacts and monitor internal AI tool use for unusual prompts or data access, since QUIETVAULT abuses locally installed AI CLI tools (pp. 14-15).
• Use CSA's MAESTRO and AI Controls Matrix to model threats against your own AI deployments and map controls (p. 16).
• Keep the fundamentals in front: attack surface reduction, least privilege and post-compromise detection, as the closing section puts it (p. 18).

Read the full document if

• you need a short briefing on GREYVIBE's campaigns, malware and AI use, and on how it fits a wider AI-equalization argument.
• you are updating your threat model to cover state-adjacent or proxy actors.
• you want a pointer list to the vendor and government reports on AI-enabled threats from 2025 and 2026 (references, pp. 19-20).

What it relies on

The paper is a secondary synthesis. Its case study rests on WithSecure Labs' May 2026 GREYVIBE report, backed by news coverage. The wider claims cite reports from NCSC, CrowdStrike, Google GTIG, Mandiant, IBM, Microsoft, OpenAI, Anthropic, CISA and others, plus policy blogs. It runs no data collection of its own. It states it was produced with AI assistance and has not had official CSA review. CSA publishes the MAESTRO, AI Controls Matrix and Zero Trust guidance that the paper recommends (p. 16).

Our read

Useful as a compact briefing that ties one disclosed threat cluster to the wider AI-misuse reporting, with a reference list to follow up. Its figures come from other publishers, so check the originals before quoting them. It also carries its own counterweight on p. 18.