Hey! I'd love to hear your thoughts, send me a voice note.

Sources:

• OWASP Top 10 — https://owasp.org/www-project-top-10-for-large-language-model-applications/

• OWASP Prompt Injection — https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html

• Check Point PuzzleMask — https://research.checkpoint.com/2026/puzzlemask-abusing-plain-prose-as-a-covert-ai-attack-vector/

• Invariant Labs MCP poisoning — https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks

• Embrace The Red SpAIware — https://embracethered.com/blog/posts/2024/chatgpt-macos-app-persistent-data-exfiltration/

• PoisonedRAG — https://www.usenix.org/conference/usenixsecurity25/presentation/zou-poisonedrag

• Authorization-First Retrieval — https://aclanthology.org/2026.trustnlp-main.15/

• Pillar Security Deadbugz — https://www.pillar.security/blog/deadbugz-currently-active-mcp-supply-chain-campaign

• APIsec Labs A2A peers — https://labs.apisec.ai/research/articles/hijacking-google-adk-malicious-a2a-peers/

• OpenAI Hugging Face report — https://openai.com/index/hugging-face-incident-and-the-road-ahead/

• Wiz Off Guard — https://www.wiz.io/blog/off-guard-breaking-litellm-from-authentication-bypass-to-cloud-compromise

• CISA KEV alert — https://www.cisa.gov/news-events/alerts/2026/09/02/cisa-adds-seven-known-exploited-vulnerabilities-catalog

• Anthropic Threat Intelligence — https://www.anthropic.com/threat-intelligence-report-september-2026

• Anthropic cybersecurity evals — https://www.anthropic.com/research/investigating-incidents-cybersecurity-evals

• Stochasticity in Agentic Evaluations — https://arxiv.org/html/2512.06710v1

• OSWorld-Verified — https://xlang.ai/blog/osworld-verified

Chapters:

00:00:00 Intro music

00:00:06 Cold open

00:01:44 Definitions

00:03:43 Why agents need a different map

00:04:29 Family 1: Injection

00:06:42 Family 2: Tool abuse

00:08:46 Family 3: Context and memory

00:10:24 Family 4: RAG and vector stores

00:12:21 Family 5: Connectors and MCP

00:13:54 Family 6: Multi-agent handoffs

00:15:51 Family 7: Auth gaps

00:17:27 Family 8: Exfiltration via tools

00:18:49 Family 9: Fake complete

00:20:14 Family 10: Nondeterminism

00:21:17 Ship gate for a small team

00:23:47 Homework

00:24:57 Close