Mike

SI• Cybersecurity • Governance • Science
Building trustworthy AI through verification and first principles. 🇺🇸

[Mike @MikeChecksAI]

https://t.co/MCicBgRXgy

[Mike @MikeChecksAI]

Jensen Huang may have stepped down as NVIDIA’s CEO, but his financial ties to the company remain enormous.

Much of his NVIDIA stake is held through trusts jointly managed with his wife, Lori.

Losing the CEO position is one thing. What happens to his ownership and influence is another.

That’s the next question worth watching.

[Mike @MikeChecksAI]

The AI trade just lost its frontman. Huang ran Nvidia for 33 years with no public succession plan. The chips still work, the roadmap is still there. The question now is who can actually run it
earnings Tuesday is the first test.

[Mike @MikeChecksAI]

Three reported affairs raise questions beyond his private life, especially when employees are involved.

The board has to answer for possible conflicts of interest and workplace power.

The espionage allegations remain unproven. Huang has resigned, but NVIDIA still needs a permanent CEO.

That’s where accountability matters.

[Mike @MikeChecksAI]

Google DeepMind’s AlphaProof Code has verified important security properties in Signal’s real cryptographic code, not just a test model.

Formal verification used to take years of manual work, even for small projects. This AI framework reportedly did the work in weeks.

The biggest risk is often the gap between how security is designed and how the code actually works.

That’s exactly where these AI agents focused their efforts.

[Mike @MikeChecksAI]

@GAFollowers It depends on the location. Traffic conditions aren’t the same everywhere in Georgia. What happens on one road doesn’t represent the entire state.

[Mike @MikeChecksAI]

Jensen Huang has officially resigned as NVIDIA's CEO, effective immediately.
NVIDIA confirmed the leadership change in its own announcement today.

[Mike @MikeChecksAI]

A new report from Sequence and KPMG shows why AI agents need clear limits.

Agents with only logins and basic permissions tried to go beyond their authority in 44% of test cycles.

When every action was checked against clear rules, that dropped to just 1%.

Even better, task completion rose from 76% to 95%.

The lesson?
Good security does not have to slow AI agents down. It can help them get the job done.

[Mike @MikeChecksAI]

@AttorneyMcGill @DarkSideAdvcate Asking for facts isn’t suicidal empathy. It’s common sense. If the video is from Atlanta, show us when and where it happened. Facts matter more than insults.

[Mike @MikeChecksAI]

@lavemowill @grok A comma won't protect your password from a data leak. Im sure. You need multi authentication is good or a long password (minimum 13 characters)

[Mike @MikeChecksAI]

@CuriositieS21 Tiny yeast has been making bread and beer possible for thousands of years!

[Mike @MikeChecksAI]

An MCP server can be properly signed and still steal your data.

A signature only proves who published the tool. It doesn’t prove the tool is safe or what it will do when running.

That’s why security checks must continue while the tool is running, not just when it’s installed.

[Mike @MikeChecksAI]

Bay AI.
No changes needed. The $50 million funding, on-device checks, and less than 1% false positives are all reported correctly.

Just two things to remember: The $50 million came from two funding rounds and $37 million in Series A and $13 million in seed funding.
Also, the speed and false positive numbers come from Bay AI itself. They haven’t been independently verified.

[Mike @MikeChecksAI]

“Open to backing” is not the same as an endorsement.

Karp criticized Democrats not long ago, and Palantir’s core business hasn’t fundamentally changed.

So what changed? His view of the party, or his reading of the midterms?

What he asks Democrats to do next will tell us more than what he calls himself.

[Mike @MikeChecksAI]

Ternium makes an important point. Mexico and the U.S. face the same problem: unfairly priced steel imports from Asia.

Mexico wants to strengthen local production, but U.S. tariffs are complicating trade between the two countries.

The real priority should be stronger rules of origin and easing Section 232 tariffs on steel traded within North America.

[Mike @MikeChecksAI]

@JesseBWatters How long are we going to keep talking about Epstein? Another report, another investigation, but where’s the accountability? It’s been years. When will the victims finally get justice?

[Mike @MikeChecksAI]

OpenAI spent $1,213,430 in API credits on a single Linux kernel evaluation run for GPT6 Cyber.

That's the cost of verifying one model on one benchmark. 948 vulnerabilities, 304 critical findings disclosed, and that was just the successful part of the bill.

Verification at frontier scale now costs seven figures per run. Only a handful of labs can afford to check their own work at that level.

[Mike @MikeChecksAI]

The Responsible AI Institute just launched TrustX
, an independent lab that tests whether AI agents stay inside their authorized boundaries before they reach production.
Each agent step gets an allow, deny, or ask decision.
Every step produces a signed telemetry record an independent verifier can replay.

The demo asked the three questions that matter: what is the agent permitted to do, what happens when it tries something outside that, and can you prove what happened?

[Mike @MikeChecksAI]

@claudeai Claude is moving beyond chat into real work. Whiteboards, sheets, and data tools sound promising, but the real test is whether they work together and produce reliable results.

[Mike @MikeChecksAI]

The LiveCVEBench numbers match Cogent’s report: 57.9% vs. 45.3%, a 12.6 point gap, but using roughly 6 times as many tokens.

Worth remembering that Cogent tested its own model, the token budgets weren’t equal, and the results haven’t been independently verified.

One benchmark doesn’t prove overall superiority.