AI — writing by Peter Bassill
AI · page 2 of 2
AI in defence, AI as a risk, and the governance in between. 23 articles
-
AI-powered threats — and the usage policy you actually need
AI is making phishing better and deepfake fraud cheaper. It is also a real risk when your own staff paste customer data into ChatGPT. Both sides of the AI coin, and a simple policy that handles both.
-
Reasoning models: what o1 changes for SOC work
Post 7 of the AI series. OpenAI's o1 launched in September with a different model shape — *think longer, reason step by step*. What this means for the SOC, where the gains are real, and where EmilyAI's purpose-specific architecture continues to win.
-
Open-source models and the on-prem option
Post 5 of the AI series. Llama 3, Mistral, Mixtral. The serious open-source LLM era arrived in 2024. What it means for security teams who do not want to send data to a hyperscaler, and how the on-prem path reads against EmilyAI's single-tin posture.
-
The hexagonal lesson: vendor agnosticism as structure
Post 4 of the AI series. Most security AI products are anchored to one vendor's platform. EmilyAI was built in 2018 with a hexagonal architecture that decouples the analyst from the SIEM matrix. Six years on, the choice is paying back in a way I did not anticipate.
-
The Copilot-for-security wave: what they actually do
Post 3 of the AI series. Microsoft Security Copilot, CrowdStrike Charlotte, SentinelOne Purple, Google Sec-PaLM — the wave of LLM-powered security assistants. What they actually do well, what they do less well, and how the framing reads against EmilyAI.
-
Deterministic inference: the property the market is losing
Post 2 of the AI series. Same input, same output, every time. A property that used to be table stakes in production systems and that LLM-based security tooling has quietly let go of. Why it matters and how EmilyAI is built to preserve it.
-
AI in cyber: the long view from 2018
Start of a six-weekly series tracking how AI in cyber security is developing through 2024 and beyond — and how each development reads against EmilyAI, the SOC analyst I have been running in production at Hedgehog since 2018.
-
INT8 quantisation, in numbers — and why INT16 is the boring choice
What "INT8-quantised inference" actually means once you do the arithmetic, why dropping from FP32 to INT8 is a cliff and dropping to INT16 isn't, and why every interesting question about putting an ML model on real silicon ends up here.