AI Strategy & Practice Is Tomorrow, and New Red-Team Findings

Tomorrow, people from across engineering, business and governance will come together at Ham Yard Hotel in London for the AI Strategy & Practice Conference.
The conference will focus on the practical decisions organisations face when moving from AI experimentation to dependable deployment.
Putting AI to work in practice
Across the day, speakers from organisations including OpenAI, PwC, Sky, Lloyds Banking Group, Bank of America, Hg, Slamcore, Ankura and BCG will explore questions such as:
How do organisations distinguish meaningful AI value from a promising demonstration?
What does it take to move an AI system into production?
How can teams maintain reliability after deployment?
How should organisations govern increasingly autonomous systems?
What evidence is enough to decide that an AI system is safe to deploy?
The programme is organised around three connected stages: identifying where AI creates value, turning strategy into operational reality, and addressing accountability, integrity and responsibility.
The focus throughout will be on practitioner experience, honest lessons and the decisions teams encounter when AI systems meet real organisations.
One track. Open discussion. No vendor pitches in the conference sessions.
Final event details
Wednesday, 21 October 2026
Ham Yard Hotel, London
Registration and breakfast begin at 09:30, followed by the conference programme from 10:30. The day will conclude with a reception and drinks from 17:00–19:00.
A small number of final complimentary tickets are available. To request one, email jovanca@safeintelligence.ai.
What happens when real-world AI agents are pushed to break their rules?
We also recently published the results of a Spec27 evaluation examining how real production conversational AI agents responded to targeted adversarial attacks.
An agent can behave correctly during routine use while still failing when someone deliberately attempts to bypass its safeguards. Our evaluation surfaced three areas that organisations deploying agents should examine carefully:
Policy violations, where agents did not follow rules they were expected to enforce.
Suspected sensitive data leakage, where agents produced information presented as protected or confidential.
Suspected system prompt disclosure, where agents generated text that appeared to reveal private instructions or configuration.
These findings require careful interpretation. Outputs resembling sensitive information do not, by themselves, confirm that genuine customer data was disclosed. However, the behaviour remains important—particularly for agents with access to sensitive records, operational tools or real-world actions.
Strong baseline results may hide serious weaknesses
One production agent passed 95% of its baseline tests.
When evaluated using a single targeted attack method, its pass rate fell to 37%, with only 7 of 19 analysed cases passing.
This illustrates two important points.
First, routine testing may not reveal weaknesses that emerge under deliberate adversarial pressure. Second, no individual red-team technique provides a complete view. Different attacks can expose different vulnerabilities in the same system.
Production-level assurance therefore requires more than a one-off safety check. Teams need to define clear behavioural boundaries, test those boundaries using varied adversarial methods and repeat the evaluation as models, tools and deployment environments change.


