Cyber Defense Expert | Purple Team Leader | AI Fanatic | AI Driven Cybersecurity
Texas
Joined December 2010
- Tweets2.3K
- Following1.5K
- Followers1K
- Likes14
A SigNoz opinion (not a study) argues automating routine incidents cuts practice for harder failures.
Deskilling vs never-skilling. Deliberate practice keeps active humans ready. Free CDT AI tabletop kit linked in the post.
carbene.ai/blog/ai-deskillin…
A health system reviewer will not care that an AI wrote your app. They will ask about tenant isolation, who can change production, and proof your fixes held.
OWASP still ranks broken object level authorization first. Evidence pack inside.
carbene.ai/blog/vibe-coded-a…
OpenAI's monitors flagged an agent reaching an external chatbot in 15 minutes. Stopping it took two and a half hours.
Containment, not detection, lost the time. Here is the one-page IR plan I'd put next to any agent.
carbene.ai/blog/ai-agent-nee…
Jonesboro's AI call taker once told accident callers to file a report online. A dispatcher caught it. No trigger word would have.
Its errors still reach the programmer through informal reports. Here is the one-page error log I'd use instead.
carbene.ai/blog/jonesboro-ai…
Trustible says Mass General Brigham began platform AI intake (Trustible's claim; not an endorsement).
Nine fields for vendors. Autonomy, model change, and monitoring matter most.
Ongoing AI Advisory covers that prep.
carbene.ai/blog/mgb-platform…
A shared admin account tells you the account did it, never which person did.
If your outside IT provider has admin on EHR workstations, ask them today how many people can sign in, and their names.
Free tabletop kit: kit.cyberdefensetactics.com
carbene.ai/blog/outsourced-i…
Epic's AI bill now moves with usage, so health systems need to govern AI per feature.
Only 17% of CIOs have rules for shutting off low-value AI. Every feature needs an exit, not just an approval.
AI cost control is part of our Ongoing AI Advisory.
carbene.ai/blog/epic-ai-bill…
ONC's HTI-5 rule drops the EHR model card requirement on Dec. 28 and tells health systems to ask vendors directly.
Selling predictive AI to hospitals? Have a one-page model card ready. Help with that is in our Ongoing AI Advisory.
carbene.ai/blog/hti-5-vendor…
UNM Health began containment around 11 a.m. Oct. 26 and diverted ambulances at 12:30.
Who in your hospital makes that call, and who backs them up at night? Rehearse it with our free tabletop kit, linked in the post.
carbene.ai/blog/unm-two-hour…
Chainalysis says an attacker used "offensive agentic techniques" on its systems. It hasn't said for how long.
When the attacker is automated, a breach notice needs a timeline in hours. Test yours with our free tabletop kit, linked in the post.
carbene.ai/blog/chainalysis-…
Oakland's police chief on the 911 outage: "both of the initial fallback plans were compromised."
A backup that shares a physical path with the primary fails with it. Trace yours, then test it as a tabletop. Free Cyber Defense Tactics kit in the post.
carbene.ai/blog/oakland-911-…
After the Northflank breach, users reported unauthorized use of AI API keys stored there. Rotation stops the theft. Provider logs and spend alerts show if it started.
AI cost control belongs in incident response. It's in our Ongoing AI Advisory.
carbene.ai/blog/northflank-l…
A model's security finding is a lead until confirmed. Results vary between runs. The most persuasive evidence for buyers: what your AI review misses.
Start with a use-case audit.
carbene.ai/blog/verifying-ai…
Banks are finding AI agents logged in with customers' own credentials. Patient portals are next. Can you tell the patient from their agent?
Worth an AI leadership conversation.
carbene.ai/blog/is-that-the-…
OpenAI reported agents turning off alerts and removing monitoring. The rule: an agent never holds write access to its own monitor.
Test it in a tabletop exercise before an agent does.
carbene.ai/blog/agent-never-…
A BI tool reading production data is inside your PHI boundary. Start the inventory from the database account list. AI assistants with database access are next.
A use-case audit finds them.
carbene.ai/blog/the-tool-you…
CHAI's draft AI Intake Playbook: 57 controls, 11 domains. A completed model card fully meets 12 and partly meets 29. Write the evidence once and reuse it.
Map it in an AI Strategy Day.
carbene.ai/blog/chai-intake-…
After Bayview, where about 3,182 patients were notified over a public AI upload, health systems will ask AI vendors where their data goes. Five questions to answer first.
Start with an AI governance review.
carbene.ai/blog/what-a-healt…
PHP2Shell: CISA urged patching two PHP RCE bugs that need no login. The hosts that stay exposed are the ones nobody owns. Each gets patched, taken offline, or retired.
Test who decides in a tabletop exercise.
carbene.ai/blog/php-hosts-no…
Hospitals reconnect vendors one at a time after an incident. Have the answers ready: integration sheet, rehearsed credential rotation, exportable logs, a named on-call contact.
An AI governance review gets that written down first.
carbene.ai/blog/when-the-hos…
