AI Architect · Sovereign AI for regulated banking · LLM red teaming before go-live · https://teapawt.net/t.co/RJq8xP5EKb
Joined April 2015
- Tweets198
- Following252
- Followers67
- Likes517
Here's the thing: the ECB gave significant banks until 31 October 2026 to send their supervisors an action plan against AI-enabled cyber threats. Annex 1, area 2: "Accelerate vulnerability and patch management at scale." Patch speed is now a board question. NOT a ticket queue.
Source: ECB Banking Supervision letter SSM-2026-0301, 7 July 2026. Annex 1 runs from attack surfaces to crisis exercises: bankingsupervision.europa.eu…
Look. Your payments agent saw "KYC passed". It did NOT run KYC. It can't verify it. The money moved anyway. SAD!
That's Borrowed Clearance.
Signed, payload-bound clearance is NEVER inherited. Can't prove the check? Then you don't have it. STOP!
Folks, what it is, how to red-team for it, and the fix:
ercin.info/borrowed-clearanc…
LOOK at footnote 3 of SR 26-2. That's the US model-risk guidance (17 Apr 2026) that REPLACED SR 11-7.
Generative and agentic AI? "Not within the scope of this guidance." OUT!
So who tests your bank's LLM? BTW, the same footnote points back to your own risk management. YOU do.
SR 26-2: federalreserve.gov/supervisi…
Footnote 3 is in the attached guidance: federalreserve.gov/supervisi…
What testing an LLM before go-live looks like: ercin.info/llm-red-teaming-b…
US banks are BLOCKING third-party AI agents from customer accounts. HUGE.
Why? Some customers hand the agent their OWN login. The bank can't tell human from bot.
Folks, that's not an AI problem. It's IDENTITY. Every agent needs its own scoped credential. Sharing yours? WRONG!
Look, the fix is boring and it works: agents get their OWN identity, scoped permissions and an audit trail. Reuters via Times of India, 23 Oct 2026: timesofindia.indiatimes.com/…
"In-country inference" is NOT "everything stays in the UAE." Read the fine print!
OpenAI's own page: GPU inference stays in region. Auth, routing, analytics and some CPU work may run outside it.
Look, your bank's data-flow diagram needs BOTH boxes. Every time!
OpenAI, data residency and inference residency (see "Processing that may occur outside the selected region"): help.openai.com/en/articles/…
Context, Emirates NBD x OpenAI, 19 Oct 2026: emiratesnbd.com/en/media-cen…
A passing LLM eval is a SNAPSHOT. Not a control.
Change the model, prompt, RAG corpus, tools, guardrails or languages? Re-run it. Every time!
btw it fits in CI: evaluate() returns GO or NO-GO, pytest fails the build on an open critical. Boring. BEAUTIFUL!
The scorer with evaluate(): huggingface.co/spaces/Ercin/…
Six re-test triggers, explained: huggingface.co/blog/Ercin/ll…
The UAE central bank's AI guidance note (11 Feb 2026) asks for "the clear and immediate ability, with human intervention, to cease use" of an AI system.
Here's the deal: if nobody has timed your kill switch, you don't have one. TESTED OR IT DOESN'T EXIST!
Section 6f, CBUAE guidance note on AI and ML: rulebook.centralbank.ae/en/r…
Where the kill switch should live (the LLM gateway): huggingface.co/blog/Ercin/so…
Pass/fail red teaming is LYING to your bank. A harmless slip and a customer-data leak get the SAME score. WRONG!
Look, severity is the signal. So I built a free scorer: 9 harm categories, 4 severity levels, 5 go/no-go gates. Open source. Nothing stored.
#LLMRedTeaming
Free, Apache-2.0, no attack prompts, no model calls: huggingface.co/spaces/Ercin/…
Views are my own.
Check out the latest article in my newsletter: AI as Actor: Why 80% of Agent Projects Fail linkedin.com/pulse/ai-actor-… via @LinkedIn
