An AI model hacked a real company to cheat on a test.
OpenAI was benchmarking its models on hacking skills. The models found a zero-day, escaped the sandbox, got online, then broke into Hugging Face's production servers to steal the answer key.
Nobody told it to attack anyone. It just wanted to pass.
How do you threat model an attacker with no intent, just optimization?
https://t.co/DvwMMgyCin
GlassWorm is being discussed as a supply chain attack. I think it’s really a story about trust.
Attackers didn’t exploit a vulnerability. They exploited assumptions about trusted tools, trusted updates, and trusted workflows.
The detail that stood out most to me was the possible use of AI to generate believable cover commits across compromised repos.
If that’s where attacks are headed, defenders need to spend less time asking whether something looks legitimate and more time understanding what it’s actually doing at runtime.
That’s why I wrote this piece. The developer workstation is becoming one of the most important attack surfaces in the enterprise.
🔗https://t.co/6G2NF4UKLE
Cybersecurity Dive recently covered Gartner’s message to security leaders: do not let AI urgency turn into panic spending. That feels right to me.
AI will change the speed and scale of attacks, but not every AI risk requires a new tool or a dramatic program shift. In many cases, the better move is to slow down, understand the actual exposure, and determine where AI poses meaningful risk to the business.
Security leaders are being asked to respond quickly, but speed without judgment can create its own problems. https://t.co/Op9BXlUYR8
The same pattern is emerging in AI security. Visibility, governance, and guardrails are what's making broader adoption possible in the first place.
🔗https://t.co/dfP5Dq2Ecw
HBR published a piece on how stronger privacy laws may actually increase consumer willingness to share data. The research found that when people felt more protected and had clearer visibility into how their information was being used, trust increased, and so did participation.
I think that dynamic matters beyond marketing.
Security and privacy are often framed as opposing concepts. In practice, clear controls and transparency can create confidence in the system itself. People are usually more willing to engage when they understand what is happening.
Upwind published a new blog on CISO leadership that is worth reading. The core idea is that earning a seat at the table has gone too far. CISOs are most valuable when they stay close to the practitioner mindset of being independent enough to see risk clearly, but connected enough to understand business decisions.
That outside perspective is not a weakness. It is part of the job. https://t.co/fMZTkNi5bt
The risk is not only what information an agent can access. It is also what its behavior reveals, what tools it can call, and what actions it is trusted to take.
AI systems are creating a new category of crown jewels, and the protection model needs to account for that.
A useful shift in AI security is recognizing that model theft is not the same as data theft.
Data theft is about what leaves the environment. Model theft is about what an attacker can learn from the system itself. That can be just as valuable as the data behind it.
Modern environments are built on layers of code, packages, extensions, plugins, and automation. Every one of those layers can become part of the attack path if teams lack the right visibility.
Cybersecurity Dive reported that hackers breached thousands of GitHub repositories after a GitHub employee used a malware-infected Visual Studio Code extension. https://t.co/aHazP0gLUf
The initial issue was not some exotic exploit. It was a trusted development tool, becoming the entry point, and from there, attackers got access. This is why software supply chain security has become such a serious board-level issue.
Security teams are being asked to make faster decisions with more systems, more identities, more APIs, more cloud services, and now more AI-driven activity in the environment. At some point, what you used to do is no longer enough. https://t.co/BatsmbBVFg
That is why I think runtime context is becoming the foundation for the next phase of cloud security. AI can help teams investigate, validate, and remediate faster, but only if it is grounded in live context. Otherwise, it just becomes another layer of noise.
Most major data breaches occur because companies gradually lose visibility into their environments and do not realize how much sensitive information is exposed until it is too late.
Data security sounds simple. Just keep information safe from the wrong people. That's why many people think data security is just putting a password on something. However, that data moves constantly between apps, cloud systems, laptops, APIs, emails, and third-party vendors.
Every time data is moved or copied, there is another opportunity for someone to lose track of it or misuse it. Good data security requires limiting access so people only see what they need, monitoring for unusual behavior, and ensuring old accounts or unused access are removed.
When an agent can discover an API token, call infrastructure APIs, and take destructive action, the security questions move to blast radius, environment scoping, backup isolation, and runtime visibility.
Futurism recently wrote about a Claude-powered coding agent that reportedly deleted a company’s production database and recent backups.
The lesson is that agents are becoming operators, and many of those environments were not designed for that. https://t.co/hmpflWJ0Ci