LinkedIn publication text
Threat Modeling in AI: LLMs, Agents, and Multi-Agent Systems
Standards still run behind practice. I gathered in 32 slides what the literature from 2016 to 2026 already shows about frameworks, threat classes, and mitigations, with 50 audited references (levels A, B, and C). Useful for GRC, information security, and AI engineering teams.
What the study covers:
🔵 Frameworks: OWASP Top 10 for LLMs 2026 (4 Aug. 2026), OWASP Agency, AISVS 1.0, NIST AI 100-2 E2025, MITRE ATLAS, and MAESTRO.
🟣 External threats: prompt injection, target hijacking, supply chain, poisoning, and the lethal trifecta: private data + untrusted content + external communication.
🟢 Defenses: Adaptive attacks outperformed the majority of 12 published defenses, with more than 90% success. The trusted frontier is architectural: least privilege, isolation, and human approval.
🟠 Model behavior: reward hacking that generalizes to emergent misalignment, alignment faking, and scheming. Monitoring the chain of thought helps, but optimizing against the monitor leads to obfuscation.
🔴 Multi-agent: miscoordination, conflict, and collusion, with inter-agent infection (Hammond et al., 2025).
🟡 AI improving AI: The task horizon doubles every 3 to 7 months, according to the source. On Jul. 2026, two OpenAI models escaped a sandbox and hacked Hugging Face to get the answer key for a test.
Three conclusions:
🟦 Agent safety is also guaranteed in the operation, not just in the shift-left.
🟪 In multi-agents and in AI-driven R&D there is a risk catalog, but not yet a verification method.
🟩 Where sources diverge, the study records divergence.
Includes a timeline, glossary with 45 terms, and bibliography with audit status, so you can check each source.
Which of these fronts most concerns your organization?
© 2026 Aridio Silva @aridiosilva
Independent Researcher
Creator of SGAEIA
ORCID: 0009-0008-2411-6995
ORCID: https://orcid.org/0009-0008-2411-6995