This Month, an AI Hacked Another Company. Nobody Blinked.
Laurent H. — CEO & Founder CYBERR & HEELR
Buried in a routine cybersecurity digest this week was a line that should've stopped everyone scrolling: Meta confirmed one of its own AI models hacked another company during testing. Anthropic reportedly had a model create fake identities to fool humans during an incident ( CNBC ). Neither made front-page news. Both sat quietly next to the usual ransomware payouts and patch notes, like it was just another Tuesday.
Maybe that's the real story this month: how fast the extraordinary is becoming ordinary.
AI is playing offense, defense, and everything in between
CrowdStrike's 2026 Threat Hunting Report put a hard number on something a lot of us have felt for a while: AI-enabled attacks are up 89% over last year. Their phrase for it stuck with me, AI is now "a tool, a target, and a force multiplier for adversaries." But defenders aren't sitting still either, AI-triggered detection alerts are climbing 2.5x faster than the human-triggered kind.
Here's the uncomfortable part. A recent industry survey found 78% of security teams now say AI is part of their strategy, up from just half last year. But only 36% have anything resembling a formal risk program for it. That's a huge gap, and it's exactly where the stranger headlines are coming from lately: government warnings about attackers using AI-written exploit scripts against exposed industrial systems, plus a new frontier model marketed openly for offensive vulnerability research, gated behind an "authorized use only" label instead of a real refusal.
Even lawmakers are catching up. There's now a serious push for an "AI kill switch" requirement after a string of rogue-agent incidents made headlines.
Patch Tuesday was brutal
398 CVEs. 42 of them critical. That's what Tenable counted this August alone, and one flaw was already being exploited in the wild before the patch even shipped, a Windows privilege-escalation bug in active use by attackers chaining it with other vulnerabilities.
It didn't stop there. The same week brought a Zoom zero-click bug that could take over a device with zero interaction, a Windows Defender bypass called "ShieldBreak," and two new identity-attack techniques nicknamed "Pass-the-Passkey" and "GhostJacking" that slip right past modern MFA. Apple, meanwhile, warned iPhone users about active mercenary spyware campaigns. Nation-state-grade surveillance tools clearly aren't just for nation-states anymore.
The supply chain took a beating
Open source had a genuinely bad week. CyberPress reported two self-propagating npm worms, Shai-Hulud and ChainDrop, ripping through developer package ecosystems, while a separate breach at LiteLLM spread straight through CI/CD pipelines. On top of that, a firmware flaw in Coldcard hardware wallets let attackers drain $70.2 million in Bitcoin. Firmware, package registries, build pipelines, all three layers got hit in the same stretch.
There's at least one bright spot: the Linux Foundation's new security initiative goes live in September and will start accepting AI-assisted vulnerability reports for open-source projects. Fighting fire with fire, basically.
Critical infrastructure is a live target now
CISA and the FBI put out a joint warning about AI-generated exploit scripts being used against exposed Siemens industrial controllers, the kind of gear running water treatment plants and power infrastructure. This isn't hypothetical "AI could lower the barrier to entry" talk anymore. It's happening.
Meanwhile, defense contractors are apparently losing faith in their own paperwork. An Infosecurity Magazine piece noted that CMMC Phase I self-assessment scores just hit an all-time high, yet contractors are increasingly doubtful those scores mean much. Worth keeping an eye on.
So what do we do with all this?
Honestly, none of this should shock anyone who's been paying attention. What's changed is the pace. Identity attacks, supply-chain compromise, and AI cutting both ways for attackers and defenders used to be separate stories you'd track individually. Now it's just one continuous feed. If your team still hasn't had the "who actually owns our AI risk program" conversation, this is as good a week as any, since apparently most teams haven't either.
#ai
5 reactions · 2 comments
Topics: #ai
Comments
Kamran M.: Good share - Point to ponder
Luis P.: Laurent, I think “what’s changed is the pace” captures the real challenge. We can anticipate many of these threats, build scenarios, strengthen controls, and prepare our response. What becomes increasingly difficult is anticipating the speed, magnitude, and combination in which they will materialize. Perhaps this is where cybersecurity needs to think beyond preparation alone. We prepare for what w