wAIser Blog
Practical insights on AI agents and AI security for companies that want to use AI responsibly.
Agent Security Audit: A Self-Assessment
Four posts, 247 papers, named exploits later, here's the practical part. Score your own agent across the trifecta, the lifecycle, your defenses, and known attack patterns, out of 16 points.
The Defense Playbook for AI Agents Is Finally Taking Shape
Zero Trust for agents, treating internal agents as insider threats, and the first certified defenses against memory poisoning. Three 2026 frameworks that share one instinct: shrink what you have to trust.
247 Papers Later: Agent Security Is a Systems Problem
A 2026 Systematization of Knowledge digests 247 agent-security papers: individual defenses work but don't compose, we measure the wrong things, and agent security is a systems-engineering discipline with an LLM at its core.
The Lethal Trifecta: Why 98% of AI Agents Are Sitting Ducks
Private data + untrusted content + external communication: 98% of production agents combine all three legs of the lethal trifecta, only 11% pass a blast-radius check. You can't prompt your way out — only architect your way out.
The Year AI Agents Became the Attack Surface
GuardFall, AutoJack, GitHub supply-chain attacks, fake bug reports with 85% hijack rates: the 2026 agent-security research reads like a pentest report. Why software with hands changes everything.
Timeless knowledge, trained AI-native: why I built FutureClass
Carnegie, Machiavelli, Le Bon: human nature hasn't changed — only the garment of words. Why reading isn't mastery, and how FutureClass makes timeless knowledge trainable with AI Practice Labs.
Snake oil for AI search: what the science says about GEO
GEO as a paid discipline is largely unproven. What the research (arXiv, the Princeton study, Fishkin's 2,961 prompts) actually shows about AI visibility, and what's just snake oil.
We Poisoned a "Security-Conscious" AI Agent in 2 Messages
No API access, no jailbreak — just 2 chat messages that made a hardened AI agent trust a credential harvester. Four experiments on Mem0 memory poisoning, with every defense tested and broken.
Loops, Not Prompts: Why the Head of Claude Code Stopped Prompting
Boris Cherny stopped prompting Claude – he writes loops. What that means, how /loop and /goal differ, and the autonomous triage loop I built myself.
Opus 4.8 & Dynamic Workflows: 7 Agents in 5 Minutes
Seven agents, a real legal-tech platform, five minutes. What it means that Claude can now turn a plan into a script and run a swarm in parallel — and why verification changes everything.
Ambient AI: Why the Next Generation Doesn't Answer – It Notices
As long as we pay AI per request, we only build one shape of it. The interesting shape is the janitor who notices on his own when something is off – and 2026 is the year it becomes economical.
Anthropic's 33-Page Blueprint for Claude Skills
An official guide showing how to teach Claude the way you work – without writing a single line of code. What's actually in it, and why skills quietly retire prompt engineering.
Did MIT solve AI hallucinations?
A new study from Cambridge promises more reliable models through calibrated reward signals. What it actually shows – and what it means for developers today.
Claude Mythos Preview: When AI Finds Zero-Days on Its Own
181 Firefox exploits, a 27-year-old OpenBSD bug, a full FreeBSD RCE with a ROP chain. Why the offense-defense balance is shifting right now.
A CLAUDE.md With 25,000 Stars
How a single Markdown file built from Karpathy's observations became the most-starred repo in the Claude Code ecosystem – and what actually holds up in practice.
Superpowers: AI Agents Like Real Teams
What Superpowers really does, what the team pattern caught in a real release with 13 agents, and when spec, plan and subagent review are overkill.
OpenClaw vs. NemoClaw: Why the Question Is Wrong
As of October 2026: what NemoClaw's sandbox does against the lethal trifecta and what it doesn't, this year's CVEs, and a hardening checklist for self-hosted agents.
How We Found the Best AI Model for Legal Analysis
Updated Oct 2026: how we evaluated AI models for legal analysis, plus an eight-step method to rerun on today's models: real test cases, reference checklists, blind pairwise judging, judge-bias checks, cost.
Legal AI Knowledge Base with Docling: Lessons from Practice
How to build a citable legal knowledge base with Docling: parsing, chunking by section and subsection, citation metadata, post-ingest quality checks and retrieval monitoring.
Multi-Agent AI Systems with CrewAI
CrewAI in production, not the tutorial: crews vs. flows, schemas that don't kill a run, invented sources, cost ceilings – lessons from my own multi-agent systems.
Got a hard
AI problem?
Whether it's building agents or breaking them — if it's interesting, I want to hear about it.
START A PROJECT ▸