Cyber News Bytes: Google's Gemini Hack, Cisco's Perfect-10 Flaw, CrowdSec's GitHub Breach

Before we get into this week’s headlines, I just published a new video breaking down exactly where the cybersecurity industry is heading.

If you want the quick rundown on what the future of cyber looks like, it's worth a watch. Now, let’s get into this week’s top cyber news stories! →

1. Google's Gemini hacked three real companies during a test, then stopped itself

Google confirmed that its Gemini AI model accessed the systems of three real companies during a May cybersecurity evaluation after a testing error unintentionally gave the model internet access. Gemini was supposed to be attacking a fictional target during a capture-the-flag exercise run by evaluation firm Irregular, but one fictional company happened to share a name with a real one, and the model treated it as fair game. In two cases, Gemini found exposed credentials sitting in public code repositories, and in a third it repeatedly guessed passwords until it broke in.

Google says the model recognized in all three cases that it had reached genuine infrastructure rather than a simulated target, and stopped on its own. Google is now the fourth major AI lab, following OpenAI, Anthropic, and Meta, to disclose a testing environment that leaked into the real world.

Why it matters: Four labs, the same root cause: telling a model it's in a simulation is not a security boundary. Only real technical controls, like egress filtering, DNS allowlists, and network isolation, actually keep a capable AI agent contained. If your organization runs any kind of AI agent evaluation or red-teaming exercise, this is the week to audit whether your "sandbox" can actually reach the internet.

Read more at Axios

2. A perfect-10 Cisco flaw was already being exploited before anyone knew it existed

Cisco disclosed CVE-2026-76460, a maximum severity 10.0 authentication bypass in Cisco Identity Services Engine, the network access control platform that decides who and what gets to connect to enterprise networks. An unauthenticated remote attacker can send a single crafted request to an API endpoint and bypass the entire web-based management interface, landing root access on the device. There is no workaround, patching is the only fix.

The most unsettling part is how Cisco found it: not through an internal audit or bug bounty, but while its own support engineers were resolving a customer's help desk ticket, meaning at least one organization had already been breached before the vulnerability was even known to exist.

Why it matters: ISE often stores the RADIUS and TACACS+ credentials that authenticate every device on a network, so a breach here can cascade into every connected switch, router, and firewall. If you run Cisco ISE, patch immediately and hunt for anomalous API calls in your logs, since attackers with root access can tamper with logs to hide their tracks.

Read more at The Hacker News

3. A departed employee's forgotten GitHub access led to 170 stolen private repositories, four months later

Security firm CrowdSec disclosed that an attacker copied roughly 170 of its private GitHub repositories back on May 22, using an OAuth token tied to an employee who had already left the company. CrowdSec had kept his GitHub access open so he could finish some remaining work, and that access became the entry point after his laptop was compromised in the unrelated TanStack npm supply chain attack, where 84 malicious package versions harvested credentials from developer machines. The stolen code sat undetected for nearly four months until it surfaced on a cybercrime forum on September 16.

CrowdSec says its production infrastructure and databases were never touched and no code was altered, but the leak did expose email addresses for 83 users and contact details for 51 potential investors from 2020. The company had actually removed the employee's account three days after the theft occurred, just not fast enough.

Why it matters: This is a textbook case of offboarding lag turning into a breach months down the line. A supply chain compromise you have no direct control over can still reach you through the weakest link in your own access management. Review your offboarding process this week and make sure departing employees lose access the same day, not "whenever they finish up."

Read more at The Hacker News

4. Three actively exploited Linux kernel flaws just got CISA's urgent patching order

CISA warned this week that attackers are actively exploiting three vulnerabilities in the Linux kernel, adding all three to its Known Exploited Vulnerabilities catalog and creating an urgent patching deadline for federal agencies. The Linux kernel underpins the vast majority of servers, cloud infrastructure, and embedded devices worldwide, making any actively exploited kernel-level flaw a high priority regardless of which specific distribution an organization runs.

Why it matters: Kernel-level vulnerabilities sit below almost every other security control an organization has in place, so exploitation at this layer can undermine defenses that otherwise look solid. If your infrastructure runs on Linux, and most of the internet does, cross-reference your kernel versions against CISA's KEV catalog this week rather than waiting for your next scheduled patch cycle.

Read more at The Hacker News

This week's stories all trace back to the same idea: a boundary that looked solid on paper, whether it was a test environment, a network appliance, an offboarding checklist, or an operating system kernel, and wasn't actually tested under real conditions. Building the hands-on skill to find and close those gaps yourself is exactly what turns theoretical knowledge into something you can actually demonstrate.

Stay sharp,
Sandra

Sandra | Cyber With Sandra