Blog
Field notes on security and AI, written by a practitioner.
Long-form writing for risk, security, and board readers. Framework-anchored, regulator-literate, and informed by what actually happens in mid-market deployments.
15 min read
Jailbroken AI as an attack tool: how attackers actually use it, and how to defend
Attackers jailbreak frontier models, rent dark LLMs like WormGPT, and bolt AI onto phishing, deepfake fraud, and agentic intrusions. The documented tradecraft, the real incidents, and the defence-in-depth controls that hold, with Rust guardrail code.
- AI agents
- Jailbreaking
- Prompt injection
12 min read
Managing cryptography in the post-quantum era: what businesses need to do now, and the attacks already hitting encryption
The post-quantum standards are finalised and ASD's Information Security Manual stops approving RSA and elliptic-curve cryptography after the end of 2030, ahead of most of the world. The operational playbook: the attacks on encryption happening now, the honest quantum timeline, and how to run the migration.
- Post-quantum
- Cryptography
- Cryptographic agility
24 min read
OpenClaw, WhatsApp and Telegram: the phone-linked AI agent threat model, the attacks already in the wild, and the gold-standard alternatives
OpenClaw lets anyone wire a personal WhatsApp or Telegram account to an AI agent in ten minutes. Bitsight found 30,000 instances exposed on the public internet in a fortnight. This is the architecture, the attacks, the config that breaks it, and the official-API pattern that holds.
- AI agents
- OpenClaw
21 min read
Long-term memory for agnostic agents: a working architecture on Bedrock AgentCore
Bedrock AgentCore gives you a managed runtime, a long-term memory store, and a deliberately framework- and model-agnostic SDK. That keeps the agent code portable while the memory plane becomes the new audit liability. Here is the architecture that uses AgentCore Memory properly, the governance controls memory-poisoning and Privacy Act obligations force on top of it, and the rollout cadence that keeps the deployment defensible.
- AWS
- Bedrock
- AgentCore
20 min read
OAuth scopes weren't built for AI agents: the delegation model that holds up under prompt injection
OAuth scopes assume a human approves once, an app does narrow work, and the trust horizon is months. AI agents break every part of that assumption. The architecture that holds is a two-principal model with short-lived delegation tokens, ReBAC for structure, ABAC for context, and per-action consent gating destructive operations. Here is the design and the rollout.
- AI agents
- Authorisation
- ReBAC
11 min read
Reporting AI risk to the board: a one-page position summary that actually works
What the board actually wants on the AI risk page is the answer to four specific questions. Most AI risk reports answer different questions. Here is the structure that lands, four worked examples by sector, and a template you can lift verbatim.
- Board reporting
- AI governance
- Risk
Stay informed
Get new posts by email.
One email a fortnight. Long-form content only, no promotional sequences. Unsubscribe at any time.
Get started
Talk to the security lead who would do the work.
A thirty-minute discovery call costs nothing. We confirm fit, scope, and timing, then issue a fixed-fee statement of work within two business days.