Social engineering in the age of AI
AI made phishing fluent and voices cloneable. What we get away with on engagements now, and the controls that still work when training is not enough.
By the engineering team
·
Aug 10, 2026
field notes · 2026
Articles from real engagements.
Patterns we see. Mistakes that keep showing up. New attack surfaces worth tracking. Written by the engineers who run the engagements, not a marketing team.
01. filter by topic
02. latest
Recent writing.
Sorted by publish date. Click a topic chip above to filter.
How to secure your AI agent before it ships
The build-side checklist that closes the findings we keep writing: least privilege, trust boundaries, output validation, memory isolation, and human gates.
ISO/IEC 42001: the AI management system, in plain terms
The first management-system standard for AI. How it maps to ISO 27001, the impact assessment it adds, and where security testing provides evidence.
The EU AI Act, for the team that actually ships
Risk tiers in practice, the dates that matter, and where the AI Act and GDPR overlap. Written for engineers, not lawyers.
AI security and penetration testing: finding the gaps in AI and agents
What an AI pentest actually tests: model, prompts, tools, and the agent loop. The failures we keep finding and the controls that hold.
What SOC 2 auditors actually accept as pentest evidence in 2026
The format has tightened. Six examples of what passes review, what gets flagged, and what your engineering team can prepare ahead of fieldwork.
Prompt injection patterns we keep finding in agentic systems
Indirect injection via documents, tool-use abuse, MCP server exposure. Notes from twelve AI security engagements over the last quarter.
The AD attack path nobody patches
ADCS misconfigurations still account for a third of the internal escalation paths we find. Why fixes stall and the four configuration changes that close most of them.
Supply-chain attacks via internal GitHub Actions
Self-hosted runners shared across repos are a quiet compromise path. A walk through one engagement and the policy changes that closed it.
Red team objectives that test what matters
Scoping a red team by asset list misses the point. Five objective patterns that produce engagements your CISO can take to the board.
Cloud IAM blast radius: the report we wish we did not have to keep writing
The same five mistakes appear in every cloud audit. Each one has a fix that takes a sprint.
No articles match that topic yet. Pick another, or choose All.
→ reference
Looking for term definitions?
Attestation letter, penetration test report, cardholder data environment, prompt injection, purple team, threat modeling, proof of concept. Plain-language definitions written by the engineers who test for them.