Secure Agentics

Everyone talks about agentic AI. No one tells you how to make it safe to ship. Here's how to build the 8-layer security harness. (Save this, I've added the link to the open source repo that ships the entire harness at the end) Security doesn't live in the prompt. It lives in the harness around the model. 1. Define the agent contract ↳ Allowed actions (the agent's declared remit) ↳ Forbidden actions (deny list, regardless of prompt) ↳ Enforcement mode (audit, human-in-the-loop, or block) 2. Capture actions and reasoning traces ↳ Tool calls intercepted before execution ↳ Reasoning traces alongside, not just outputs ↳ Session correlation across multi-turn runs 3. Scrub PII before it leaves ↳ Client-side regex redaction (emails, phones, IDs) ↳ Server-side LLM sweep (names, contextual identifiers) 4. Analyze the reasoning trace ↳ Intent caught in the reasoning before the action runs ↳ Scope drift flagged from the chain of thought ↳ ~35% detection lift over action-only monitoring 5. Harden the analyser ↳ Sandboxed analyser model (no tools, no MCP, no internet) ↳ Inputs spotlighted as untrusted data (separated from system instructions) ↳ Structured-only outputs (no free text, no prompt leak) 6. Tier the verdict by severity ↳ Multi-tier severity scoring (low risk through critical) ↳ Per-agent action threshold (block, alert, or pass-through) 7. Choose the control mode ↳ Audit mode (log the verdict, let the action run) ↳ Human-in-the-loop mode (pause the action, route for approval) ↳ Block mode (pre-action automatic refusal based on agent scope definition) 8. Push alerts to engineering channels ↳ Severity-gated firing (only above your threshold) ↳ Channel routing (Slack, Discord, or existing alerting stack) ↳ Alert payload (verdict, reasoning trace, blocked action) A safe production agent is not just a model with a system prompt. It's a system with a contract, capture, redaction, reasoned evaluation, hardening, verdict tiers, control, and alerting, all working together. ___ I found a completely open-source SDK that does all of this in just a two-line integration that wraps your agent. Check out the GitHub repo and star it so you can come back to it later: https://lnkd.in/efMuFtgy ♻️ Repost if this is useful to an engineer shipping agents.

We've open-sourced what we believe is the most capable AI agent security monitor on the market. Yes, more capable than the ones you're paying for. Adrian is live on Product Hunt today 🩵 https://lnkd.in/euiNSG7u

Last week we launched our first product, an open-source runtime security toolkit to monitor agents that think (https://lnkd.in/eDBQX3fj). The community response has been epic - THANK YOU! Here is the story in a few numbers: - 713 downloads (pip) - 261 git clones - 46 GitHub stars - 10 forks, 4 Issues & 2 PRs raised by the community (some swag going to DHRIT PATEL as our first community contributor!) - 90 new members in our discord community (https://lnkd.in/egY4zmhU) All this in just 6 days. That is the power of community. That is the power of open source. And we're just getting started!

We are thrilled to announce that Secure Agentics will be attending RSAC 2026 in San Francisco as part of the Laboratory for AI Security Research (LASR) Delegation, delivered by Plexal. LASR is a collaboration between the public and private sectors in the UK to bring together the best minds in AI security, conducting cutting-edge research at the intersection of AI and cyber security. We are honoured to be one of just a handful of start-ups selected to be flown out to RSA, meet with US stakeholders, and push the frontier of AI security research. #AI #security #research #RSAC #RSAC2026 #LASR #Plexal

We are incredibly excited to have our very first media appearance being on BBC News - start as we mean to go on! Be the first to hear about what we are building: https://lnkd.in/etni9FdJ #AI #security

We're proud to be attending our second Cyber Innovation Competition in 3 weeks!

We're excited and honoured to be sharing the stage tomorrow with 10 other top cyber innovators, pitching to a room full of cyber leaders from industry, Government and academia! This will be the first time we're going to share publicly what we've been working on, and we couldn't think of a better stage - catch you tomorrow.