How to Detect and Hunt Attacks Across Your AI Ecosystem | Thomas Roccia | S2 E2

How to Detect and Hunt Attacks Across Your AI Ecosystem | Thomas Roccia | S2 E2

by Eva Benn

Applied Sciences & Engineering Computing & Cybersecurity UComputing and Information TechnologyURComputer security
🎙 Eva Benn 👥 101K 📅 July 27, 2026 ⏱ 38 min 👁 22K 📄 expert opinion 🧭 2026-08-16
Available in: English (current) Français

Keywords

AI threat intelligenceprompt injectionNovaindicators of prompt compromiseAI agents

Summary

In this episode of Security Mondays, host Eva Benn interviews Thomas Roccia, an AI threat researcher and founder of SecurityBreak. The conversation centers on the evolving landscape of AI threat intelligence, emphasizing that traditional threat intelligence is insufficient for the new attack surface introduced by AI systems. Roccia explains that AI ecosystems, including agents, models, and MCP servers, create new vulnerabilities that attackers exploit. He introduces the concept of ‘indicators of prompt compromise’ (IPCs) as a broader category than prompt injection, covering any unauthorized prompt activity. He demonstrates Nova, an open-source prompt pattern matching tool that allows defenders to create rules to detect malicious prompts, similar to Yara for malware. Roccia outlines three major ways attackers abuse AI: software supply chain attacks via malicious packages, AI SEO poisoning to trick users into executing malicious commands, and leveraging AI agents for automated attacks. He also critiques the Anthropic report on autonomous espionage as lacking actionable threat intelligence. The episode concludes with advice for security leaders to gain visibility and observability into their AI systems, using tools like SecurityBreak.

177 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides valuable insights into AI threat intelligence, a relatively new and critical field. Roccia’s arguments are well-structured and supported by his practical experience and the tools he has developed. He effectively explains complex concepts like prompt injection and IPCs, making them accessible to a technical audience. The demonstration of Nova adds concrete value, showing how defenders can implement detection mechanisms. The discussion of real-world attack vectors, such as AI SEO poisoning and supply chain attacks, grounds the conversation in practical threats. However, some claims, such as the prevalence of AI-driven attacks, are mentioned without specific data or citations within the video, relying on general references to reports. Overall, the argumentation is solid, but the lack of detailed evidence for some assertions slightly weakens the overall rigor.

Scientific Rigor, Source Quality, Title Accuracy

The video maintains a high level of scientific rigor, with Roccia referencing several reports and tools, including the CrowdStrike 2026 Global Threat Report, Check Point’s threat intelligence report, and his own projects like Nova and PromptIntel. The sources cited in the description are relevant and credible, though the video itself does not always provide specific citations for claims made. The title accurately reflects the content, which focuses on detecting and hunting attacks across AI ecosystems. The discussion is well-organized and stays on topic, with the demo providing practical evidence. However, the host’s promotional tone and the mention of sponsorships (though not named) could introduce bias, but this does not significantly detract from the technical content.

258 words

Title / Content Match

The title accurately reflects the content, which focuses on detecting and hunting attacks across AI ecosystems.

Quality & Reliability

8/10

The video features a recognized AI threat researcher with extensive experience, and the discussion is grounded in practical tools and real-world examples. However, some claims lack specific citations within the video, and the promotional nature of the host's platform may introduce bias.

Chapters

Cited Sources

  • SecurityBreak Blog — Thomas Roccia's blog where he publishes research on AI threat intelligence.
  • Introducing PromptIntel and Indicators of Prompt Compromise — Article introducing the concept of indicators of prompt compromise and the PromptIntel database.
  • Malware Reverse Engineering is no longer a human problem — Article discussing the use of AI agents in malware reverse engineering.
  • NOVA Documentation — Documentation for the NOVA prompt pattern matching tool.
  • NOVA — Official website for NOVA, an open-source prompt pattern matching system.
  • PromptIntel — Free adversarial prompt database.
  • Unprotect Project — Open database of malware evasion techniques.
  • CrowdStrike 2026 Global Threat Report — Report referenced in the discussion on AI-driven attacks.
  • PwC on the AI-orchestrated GTG-1002 campaign — Report on an AI-orchestrated cyberattack campaign.
  • Check Point threat intelligence report, June 29, 2026 — Report mentioning 26,000 hijacked AI agents.

Concurring Sources

Dissenting Sources

  • Anthropic's report on autonomous espionage — Roccia criticizes this report for lacking actionable threat intelligence, describing it as a marketing paper.

External References

Contribution & Novelties

The video offers a novel perspective on AI threat intelligence, introducing the concept of indicators of prompt compromise (IPCs) as a broader category than prompt injection. Thomas Roccia’s NOVA tool is presented as a pioneering solution for prompt pattern matching, enabling defenders to detect malicious prompts in a way analogous to Yara for malware. The discussion also highlights the importance of visibility and observability in AI ecosystems, a point often overlooked. The practical demonstration of NOVA adds tangible value, showing how organizations can implement detection mechanisms. The video also critiques existing reports, such as the Anthropic autonomous espionage report, for lacking actionable threat intelligence, which is a valuable critical perspective.

Pour aller plus loin :

155 words

Radar Profile

The radar profile shows high scores in quantity and quality of information, reflecting the depth of the discussion and the credibility of the guest. The technical level is also high, with detailed explanations of tools and concepts. The overall reliability is strong, though the promotional tone of the host slightly lowers the score. The profile suggests a well-rounded, informative video suitable for cybersecurity professionals.

Reliability 8/10

💬 Sur les 0 commentaires analysés, aucune tendance n'est disponible.