Blog
Notes from the field
What we run into on real engagements: Microsoft 365 tenants, AI agents, and the security questions that come with them.
-
Your AI is only as good as the data it can reach
AI answers are only as reliable as the sources behind them. The best data lives in APIs, registries, publishers, and your own files. Here is how to connect it.
Read the post -
PHP emergency patch: two critical RCEs land Wednesday
PHP ships emergency fixes for two critical remote code execution bugs on October 21 at 14:00 UTC. PHP 8 is expected to get patched. Here is the move.
Read the post -
Fast AI screens are tempting. Regulated data changes the math
A cheap classifier gate that keeps sensitive data away from LLMs only works if the gate itself is safe for that data. What to demand before wiring one in.
Read the post -
The email your AI classifies is also instructions
Categorization engines read untrusted email and web text before anyone else does. Jev's typed outputs limit the damage, but only if you wire them in first.
Read the post -
Memory poisoning: yesterday's infection is tomorrow's truth
A corrupted entry in your AI's knowledge base reaches every future session. OWASP lists memory poisoning as a top agentic threat of 2026. How to contain it.
Read the post -
What Claude actually does about prompt injection and poisoned memory
Anthropic's safeguards against prompt injection and memory poisoning, in plain terms. What holds, what does not, and what to check before trusting an agent.
Read the post -
Customer Lockbox doesn't cover Copilot Studio audit logs
Customer Lockbox does not cover Microsoft Copilot Studio audit telemetry or Agent 365 events. Here is what that gap means for regulated tenants.
Read the post -
Five Eyes agentic AI guidance: the SMB summary
Australia, the US, Canada, New Zealand, and the UK published joint agentic AI guidance in May 2026. The SMB-readable version, without the government prose.
Read the post -
The email your AI reads is an instruction
Indirect prompt injection turns everyday content into attack surface. OWASP ranks it the top LLM risk for 2026. Here is how it works and what actually holds.
Read the post -
The lethal trifecta: when AI agents leak data
Security researcher Simon Willison named the three capabilities that turn an AI agent into a data leak. If your agent has all three, one email is all it takes.
Read the post -
NIST AI RMF vs ISO 42001: what a small business needs
Two AI governance frameworks, one voluntary and one certifiable. What each actually is, how they fit together, and what makes sense for a small business.
Read the post -
Seven questions to ask any AI vendor
Before an AI vendor touches your data, make them answer these seven questions. Based on the due diligence discipline regulators expect from finance.
Read the post -
Your AI agent is a privileged user
Every AI agent holds permissions, and those permissions, not the model, decide the damage a compromised agent can do. What the Five Eyes agencies say to do.
Read the post