Why this, why now
Agents already act in delivery pipelines and production. Few organisations can show who approved what an agent did.
The numbers
- 54%
- of organisations have no defined way to limit what agents can access, in a Gartner survey of 297 security leaders, Q2 2026. Gartner's advice is to govern agents "based on action privileges". Gartner, 30 September 2026
- 6.6%
- of developers trust AI output for important decisions, while 26.2% use agents at work. Stack Overflow 2026 developer survey
- Over 40%
- of agentic projects are forecast to be cancelled by 2027, partly for "inadequate risk controls". Gartner
The gap
OWASP names the risks in its Top 10 for Agentic Applications. CoSAI, NIST and CSA design the architecture and controls. This forum is where engineering leaders bring real pipeline evidence and agree what proves who approved an agent's action.
Where we sit
Technical community is the "how we proved it in production" layer under their "what can go wrong" and "how to design it" layers. We cite OWASP ASI codes and CoSAI papers rather than publish our own risk lists, and push accepted patterns upstream as case studies.