Black Hat · November 17, 2026

PHANTOM-B: turn LLM-specific threats into concrete design-review questions

PHANTOM-B: turn LLM-specific threats into concrete design-review questions video thumbnail
Why it matters

Adam Shostack's Black Hat presentation is accompanied by PHANTOM-B, a published threat-modeling method for teams integrating LLMs. Its eight prompts cover prompt injection, hallucination, anthropomorphization, non-explainability, training issues, over-reliance, missing security engineering and biases. The whitepaper starts from an architecture diagram and asks how each concern could manifest in each LLM-facing component: what data can influence it, what decisions it controls, where information can leave and which outputs require justification. PHANTOM-B supplements whole-system methods such as STRIDE; ordinary identity, network and software-security analysis still applies to the surrounding components. The author presents the mnemonic as prompts for discovering specific threats, rather than an exhaustive classification system. Its scope is LLM callers, not model-training organizations, and it deliberately does not prescribe a catalog of defenses. The useful artifact is a set of system-specific failure scenarios that a team can assign and test.

My takeaway: Bring a data-flow diagram to the next design review, identify the LLM components and work through the eight prompts. For each relevant scenario, record the exposed asset, responsible owner, chosen control and a test that would show whether the control works.
Keep exploring

More curated notes connected through AI Red Teaming and Agent Security.

OpenAI News · framework

Pacing model development in an era of cyber-critical capabilities

OpenAI says preliminary evidence that Astra may meet its Critical cybersecurity threshold led it to pause frontier reinforcement-learning work for two weeks and keep its largest planned run on hold. New safeguards include stronger workload and network isolation, continuous boundary testing, token-level monitoring that escalates suspicious tool activity, and broader alignment checks for deception, reward hacking, and unauthorized access.

OWASP GenAI Security Project · guide

OWASP Top 10 for Agentic Applications for 2026

OWASP's community guide organizes agentic-system risk into ten categories, including goal hijacking, tool misuse, identity and privilege abuse, memory poisoning, insecure inter-agent communication, cascading failures, and rogue-agent behavior. It provides a shared taxonomy and mitigation starting point rather than a certification checklist or evidence that a deployed system is secure.

OpenAI News · framework

A blueprint for democratic governance of frontier AI

OpenAI proposes a three-part U.S. frontier-AI governance model: harmonize emerging state safety laws into a federal baseline, strengthen CAISI as an evaluation and standards institution, and coordinate a broader resilience program. Proposed controls include severe-risk evaluations, transparency reports, independent audits, safety-incident reporting, model-weight security, whistleblower protection, and periodic technical assessments.