We lead in security of Cloud, AI and Zero Trust. Follow our research, education, certification and events.
Global
Joined March 2009
- Tweets17.5K
- Following272
- Followers19K
- Likes1.8K
Most teams secure AI agents the way they secure apps: lock down the model, review the prompts, call it done.
But agents don't just generate text. They hold credentials, call tools, and hand work to other agents. The risk lives in how all of that is governed.
The CSAI Foundation launched in March 2026 to build security for that layer. Start here: csai.foundation/
#AgenticAI
1
2
178
Four days left to grab a seat.
The Verifiable Agent Summit (San Francisco, Nov 16) is a free, in-person look at verifying what agents do, what they're allowed to do, and whether you can trust the answer.
events.zoom.us/ev/AtEpq1n7kf…
1
170
Notion's vendor promised zero data retention. Notion's AI chat transcripts were still sitting at rest in a dataset at Fireworks AI when attackers got in.
Notion wasn't breached. The way in ran through a cloud provider hosting Fireworks' R&D machines. Customers heard more than two weeks after the intrusion.
Ask your SaaS vendors which inference provider handles each AI feature. Then ask which processing paths the ZDR claim covers.
labs.cloudsecurityalliance.o…
#AISecurity
1
1
199
CISO Daily Briefing: The wfx campaign hijacked 4,837 GitHub accounts to harvest CI secrets and rewrite AWS trust policies; rotating keys isn't enough, hunt the artifacts. Apache mod_http2 CVE-2026-119618 (CVSS 9.8 per Red Hat) is fixed in 2.4.71. EU AI sandbox rules give no safe harbor for third-party harm. Notion's Fireworks leak shows zero-retention promises need verifying per inference path.
labs.cloudsecurityalliance.o…
The conversation nobody's having: offboarding.
When an employee leaves, access gets revoked. When an agent's project gets shelved, its credentials just... keep working. Quietly. Forever.
Non-human identities already outnumber people, and agents will multiply them faster than anyone is tracking. Static inventories won't catch what a live agent actually does at runtime.
CSAI Foundation is building for this: csai.foundation/ #AgenticAI
2
3
288
Quick challenge: if your cloud provider had an outage or a dispute tomorrow, could you say where your most important data lives, who holds the encryption keys, and how you'd get it back?
If any part of that answer is a shrug, you've found a gap worth closing. CCSK covers it from the ground up: cloudsecurityalliance.org/ed…
1
1
277
Identity is the control plane for AI. Who an agent is, what it's allowed to touch, and how you revoke it.
That's the ground the AI Security Summit 2026: Zero Trust covers on Nov 18-19. Free and virtual, so join from anywhere.
events.zoom.us/ev/AiLBTLhQat…
6
380
Vendor security questionnaires: 60 questions, three weeks of back-and-forth, and an answer that's already outdated when it arrives.
Before you send one, check whether the vendor is already in the STAR Registry. Many providers publish their assessment there for anyone to read, so you can start from real answers instead of a blank form.
Search it here: cloudsecurityalliance.org/st…
#CloudSecurity
Trust in an AI agent shouldn't be a feeling. It should be something you can check.
Verifiable Agent Summit. San Francisco. November 16. Free and in person, with limited seating.
events.zoom.us/ev/AtEpq1n7kf…
227
Editing 7 neurons out of 12,000+ was enough to turn an 8B agent model into a sleeper backdoor. No retraining.
Microsoft Research's "Lazarus" fired an attacker-chosen command in 98.8%+ of triggered sessions and in none of the clean ones. A 10-neuron demo ran download-and-execute as root.
If you pull third-party checkpoints, diff against the declared base model first. A scanner can't prove a model is safe.
labs.cloudsecurityalliance.o…
#AISecurity
234
CISO Daily Briefing: An SGLang RCE let an attacker watch "confidential" TEE inference at NEAR AI; attestation shows which build runs, not that it's safe. Claroc found 94.7% of Grok Build apps open without sign-in. Australia accepts 18 recommendations toward an AI safety regulator and incident reporting. Stolen AI keys: check who pays.
labs.cloudsecurityalliance.o…
1
3
339
The uncomfortable math of agent adoption: a team can wire up a new agent in a day. Security review of that agent takes a quarter.
So the review gets skipped, or the agent ships first and gets "assessed" later. Either way, capability outruns control.
The fix isn't slower deployment. It's a control plane that exists before the agent does. That's the work happening at the CSAI Foundation: csai.foundation/
#AgenticAI
Letting an AI assistant connect to every internal system "just to be helpful" is like hiring a brilliant new contractor and handing them a master copy of the filing room, the payroll cabinet, and the CEO's inbox on day one. Nobody doubts the talent. The problem is that nobody decided what they should actually see.
Good AI security starts with scoping, not hoping. TAISE covers how: cloudsecurityalliance.org/ed…
"Never trust, always verify" gets harder when the thing requesting access is an agent acting at machine speed.
Nov 18-19: AI Security Summit 2026: Zero Trust. Free, virtual, two days of practical strategy for securing AI.
events.zoom.us/ev/AiLBTLhQat…
219
Shadow AI is already in your org. Someone wired a chatbot to customer data last quarter, and nobody wrote down which risks it carries.
The AI Controls Matrix gives you a place to start: 243 control objectives across 18 domains, covering model security, data governance, and supply chain. Map one AI system to it this week and see which gaps show up.
It won a 2026 CSO Award.
cloudsecurityalliance.org/ar…
#AISecurity
1
1
273
One week out. The Verifiable Agent Summit comes to San Francisco on Monday, Nov 16: a free, in-person day on how to verify AI agent actions, authority, and trust.
Meet the security leaders and practitioners working on it. Space is limited.
events.zoom.us/ev/AtEpq1n7kf…
1
2
309
Your AI coding agent may be installing abandoned, vulnerable MCP servers because it remembers them from training data.
Glow Labs found that in sessions where a deprecated server was installed or suggested, the agent supplied the package name itself in 98% of cases. In one, it saw npm's deprecation warning and installed anyway.
Those servers never appear in a manifest, so SCA and EDR miss them. Allowlist at execution time.
labs.cloudsecurityalliance.o…
#MCP #AISecurity
CISO Daily Briefing: GPUStack CVE-2026-102895 (CVSS 9.8) is in KEV, with a federal deadline of Nov 12; the fix is 2.2.2. Qubole's own AWS credentials were used to assume dormant customer roles, so review CloudTrail back to Oct 1. ENISA's NIS2 report scores EU cybersecurity maturity at 71.53, with CSIRT maturity weakest. N-able took its N-central hosted service offline after 33 flaws were published.
labs.cloudsecurityalliance.o…
4
370
Zero Trust was built for users and devices. Now the "user" is an AI agent with its own credentials and its own goals.
The CSA AI Security Summit 2026: Zero Trust is a free, two-day virtual event on Nov 18-19 about securing identity, governing AI, and protecting the autonomous enterprise.
Register: events.zoom.us/ev/AiLBTLhQat…
1
8
515
AI agents are already taking actions in production. Can you prove which agent did what, and under whose authority?
That's the question at the Verifiable Agent Summit, a free, in-person event in San Francisco on November 16. Space is limited.
Register: events.zoom.us/ev/AtEpq1n7kf…