An AI Data Enclave needs permitted model destinations before cheaper answers. Microsoft says Foundry failover stays within your configured model subset. My test: where does the whole conversation go when the first model is unavailable?
roblinton.com/notes/approve-…
1
AI search still needs the file permissions. Microsoft's SharePoint indexed source is now GA, with supported access controls. In a Data Enclave, your AI should only retrieve what you can read. Test two users with different access.
roblinton.com/notes/ai-searc…
13
An AI data enclave should let staff and their AI use information under the same access rules. My MilCIS 2026 talk covered the components, conversation classification and approved model routes. (The text box is not a declassification service.)
roblinton.com/writing/ai-dat…
4
Google's cEKM preview keeps root keys outside Google Cloud while key operations run in a hardware-attested confidential environment. Worth testing, including revocation and outage behaviour.
roblinton.com/notes/google-c…
6
Anthropic is shortening disclosure windows for new reports it sends to vendors from 1 January 2027: 30 days for high/critical flaws, 45 for others, with exceptions. Budget for the people who test and ship fixes, too.
roblinton.com/notes/ai-bugs-…
3
Check the serving stack before blaming the model. Upbound's 2 November account reports 30% versus 100% on a copy test with the same weights and GPUs. One team's diagnostic, not an engine ranking.
roblinton.com/notes/check-th…
9
Australia needs more than AI-written emails. For the NVIDIA poster paper I needed 8 H100s; the only capacity I could find was in Japan. Ordinary GPU instances. No confidential computing required. We need capacity local teams can actually use.
roblinton.com/writing/austra…
6
Notion's Fireworks update says exposed chat transcripts contained API keys. Removing a name does not remove the secrets. My read: the analytics copy needs its own access and retention review.
roblinton.com/notes/notion-a…
4
Europe wants AI in its own voice. My read from GTC Berlin: open models, local knowledge and people who want to build it themselves. It takes skills and compute too. The prawns remain under discussion.
roblinton.com/writing/gtc-be…
6
Off to GTC Berlin next week, Roman helmet and all. Last time that went badly: Varus lost three legions to Arminius, an insider with perfect Roman credentials. An AI agent with a valid login is an insider too. Check every read it makes at the data.
roblinton.com/writing/gtc-be…
37
Joint advisory AA26-281A: thieves "restricted access to the exfiltrated data to only IP addresses from Xiamen, China." Once stolen, the mail finally got an allowlist and a jurisdiction. "It's in the cloud somewhere" isn't a residency policy.
roblinton.com/notes/stolen-e…
24
Denmark's biggest CPR breach (8.8 million registered people) was caught by neither alarms nor spot checks but by "the unusually expensive bill", per DR. The one intrusion detection system nobody switches off: billing. Volume is a signal.
roblinton.com/notes/denmark-…
11
AWS will take its European Sovereign Cloud off the global backbone for several hours on 24 Oct. A real test, not a slide. But it proves the network can stand alone, not who can be made to hand the data over. That's still about who holds the keys.
roblinton.com/notes/aws-sove…
7
Denmark's DTU: attackers got into its identity system holding ~40,000 active and ~160,000 former users, with data back to 2003. Former users' CPR numbers were still there. An identity system holding leavers' national IDs is an archive with a login page.
roblinton.com/notes/dtu-brea…
18
None shall pass! Breach statements follow the Black Knight's script: 'tis but a scratch, then a flesh wound, then a draw. None of them lying, none able to see the arm on the ground. Log every read at the data and the first number gets closer to the last.
roblinton.com/writing/none-s…
17
An AI agent treats "no" as an obstacle. Three September cases (a Medicare portal, OpenAI's DNS escape, Spain's first agent breach notice) point one way: the only no that holds sits at the data. Published guidance covers three of five properties. roblinton.com/writing/agent-…
24
OpenAI's agent got out of its sandbox through DNS. Alarm at 12 minutes, a human at 15, run killed 2 h 44 min later; tool-use on their most capable models now paused. DNS is the hole nearly every locked-down network has. The human was the kill switch. roblinton.com/notes/agent-go…
1
1
59
Medicare's stats portal said no. OpenAI's agent went around it (18 June; found in August; an email to a public inbox on 10 Sep; the PM on 24 Sep). My read: the only "no" that holds is the one with nothing behind it. roblinton.com/notes/medicare…
19
Five years ago a Taiwanese utility put its file sharing on premises, keys and all, as the world moved to cloud. Then the CLOUD Act, the EU Data Act, AWS's EU-only cloud and Anthropic's customer-held keys landed on one design: the owner rules the data. roblinton.com/writing/where-…
13
In May, four governments (ASD's ACSC + Canada, NZ, UK) told CISOs to govern what their AI reads: least privilege for AI tools, policies on what they may access, auditable actions, and a vendor question list asking where the data goes. My close read: roblinton.com/writing/asd-ai…
25
