tags/ai-security
63 posts tagged #ai-security · all tags
- Conclusion: what you built, and what's next
- The app-layer red team: GORED
- Make it idiomatic: the repo refactor
- Agent memory architecture
- Context compaction and summarization injection
- Exactly-once side effects
- Durable execution and resume-verify
- Output authenticity
- Data and privacy governance
- Incident response and the kill switch
- Admin audit and one-click rollback
- The agent gets a wallet
- A governed control plane and operator roles
- The UI and chat
- Calendar CRUD and export
- Action links
- The directory and contacts
- My week at a glance
- Tasks and reminders
- A full-chain attack, contained
- ADD on an app that is not this one
- A live red-team against the real model
- Honest eval methodology
- From pattern to library
- The one-command scorecard
- Hallucinated dates
- Conflicts and mangled data
- Memory poisoning and lifecycle
- Cross-tenant leak
- Index poisoning
- Answering what is on this month
- Defeating the lethal trifecta
- Model artifacts
- A pinned tool catalog
- Agent-to-agent hand-off
- The dual-LLM brain
- No shell
- Egress-deny and the one way out
- The detonation chamber
- The problem with running code
- Circuit breakers and runaway loops
- The human gate
- Least privilege
- A web-search tool that hits a local server
- The tracking pixel
- The calendar-field exfil
- Scrubbing secrets and PII at ingest
- System-prompt leakage
- The email is data, not orders
- When the email gives orders
- One trace id, and the command stream
- A gateway you cannot route around
- A log you can trust
- Attacking the baseline, all red
- The eval harness: F1 as a receipt
- The watched folder
- Inert output and the human accept
- Dates the model never decides
- One shape for everything
- Attack-Driven Design, shown once
- Set up your lab: from a bare Mac to a running agent
- The plan: a local email-to-calendar agent, built to be attacked
- Start here: what this book builds