Guardians of Agents

Syllabus

Curriculum

The series moves from why agents are dangerous, through real incidents and the controls that answer them, to the industry, the research frontier and what comes next. One issue is published each Monday.

Why Agents Are Dangerous

What makes an agent different from a chatbot, and the risk maps defenders use.

  1. 01Why AI Agents Are Dangerous: Autonomy, Access and Untrusted InputAn AI agent is a language model with tools, memory and permission to act without asking. That combination turns every web page and email it reads into a possible command. Here is why, and what this series will cover.Published
  2. 02The Risk Map for AI Agents: OWASP ASI01-ASI10, ATLAS and MAESTROTen OWASP agentic risks with a real example each, the LLM risks that still matter under an agent, MITRE ATLAS's new agent techniques, and a way to pick the right framework without drowning in them.Coming Sep 28

Incidents & Aftermath

What has gone wrong with AI agents around the world, and what changed afterwards.

  1. 03When Coding Agents Go Wrong: The 2025-2026 Incident FileFrom Replit's deleted database to a GitHub issue title that poisoned Cline's releases: a worldwide file of coding-agent incidents, what caused each one, and what changed afterward.Coming Oct 5
  2. 04From Inbox to Espionage: Enterprise, Browser and Attacker Agent IncidentsEchoLeak, ForcedLeak, GeminiJack, hijacked browser agents, the OpenClaw scramble and state hackers running agents of their own: a worldwide file of agent incidents outside the IDE.Coming Oct 12
  3. 05The Aftermath: What Two Years of Agent Incidents ChangedAgent incidents forced patches, new bounty rules, CVEs for prompt injection, a court ruling, insurance exclusions and browser bans. Here is what changed, what held, and the patterns behind fifteen cases.Coming Oct 19
  4. 06When Agents Act on Their Own: Scheming, Sabotage and ShutdownLab tests show models editing shutdown scripts, blackmailing and gaming their graders. Here is what the primary research found, what it did not find, and how to design agents that stay safe either way.Coming Oct 26

Control & Guardrails

Kill switches, guardrails, identity and protocols, the machinery for keeping agents in bounds.

  1. 07The Agent Kill Switch: Building a Stop Button That WorksLaws now ask for a stop button, but an agent is a loop plus credentials, sandboxes and side effects. Here is a layered kill-switch architecture, what vendors ship, and where stopping fails.Coming Nov 2
  2. 08Agent Guardrails: What Each Layer Catches, and What Gets ThroughClassifiers, monitors, policy engines, human approval and secure-by-design patterns each stop some agent failures. Here is what the measurements say about each one, how attackers get past them, and how to layer them.Coming Nov 9
  3. 09Identity for AI Agents: Who Is Acting, and on Whose Authority?Agents act for people, for themselves and for other agents, often with long-lived keys. How delegation, MCP authorization, agent identities and short-lived credentials keep an agent from becoming a confused deputy.Coming Nov 16
  4. 10Agents Talking to Agents: Securing MCP, A2A and Multi-Agent SystemsMCP connects agents to tools and A2A connects agents to each other. Both secure the pipe, not the words inside it. Here is how protocol and multi-agent attacks work, and the controls that stop one bad hop from becoming a cascade.Coming Nov 23

Red Teaming Agents

How to attack your own agents before someone else does.

  1. 11How to Red Team an AI Agent: Goals, Harnesses, Benchmarks, MetricsRed teaming an agent means judging what it does, not what it says. A practical method: threat model, canary goals, a test harness, benchmarks and tools like AgentDojo and PyRIT, adaptive attacks, and metrics that survive scrutiny.Coming Nov 30

Industry & Players

How vendors, standards bodies, governments and startups are responding.

  1. 12Who Writes the Rules for AI Agents? Standards, Labs and GovernmentsOWASP, CoSAI, NIST, the frontier labs and governments from Washington to Beijing have published guidance on AI agents since 2025. Who said what, where they agree, and what is still missing.Coming Dec 7
  2. 13The Players in Agent Security: Platforms, Incumbents and a Buyout WaveBetween 2024 and 2026, security incumbents bought more than twenty AI security startups while cloud platforms built agent controls into their products. A neutral map of who does what, and what it means for buyers.Coming Dec 14

Research & Future

The academic frontier, promising technologies, and what comes next.

  1. SRWe Mapped 1,472 Papers on AI Agent Security. Here Is What Works.We classified 1,472 papers on attacking, defending and controlling AI agents. Most defenses look strong in their own papers and weak under adaptive attack. The designs that hold up best are the least tested.Coming Sep 24
  2. 14Eight Research Branches Shaping AI Agent SecurityAgent security research has split into eight branches, from prompt-injection robustness to AI control and human factors. Here are the questions each one asks, the papers that define it and the groups to follow.Coming Dec 21
  3. 15What Might Work: Promising Technologies for Securing AI AgentsPlan/data separation, microVM sandboxes, confidential computing, agent tracing, signed artifacts and policy engines each close part of the agent attack surface. Here is what each one does, and how mature it is.Coming Dec 28
  4. 16AI Agent Security to 2030: Forecasts, Deadlines and Our PredictionsWhat analysts, the International AI Safety Report and METR's task-length data say about agents through 2030, the legal deadlines already set, our own labeled predictions, and a recap of the whole series.Coming Jan 4