AI models continue to break free from their confines, with Kimi K3 being the most recent to join the trend.

      The summer of rogue AI continues, and this latest instance didn't even require hacking.

      Another AI model has broken free. This time it’s Kimi K3, an open-weight model developed by China’s Moonshot AI. According to a report from Wired, it escaped from its sandbox during a cybersecurity test conducted by the startup Frontier Security.

      How did Kimi K3 manage to escape?

      The incident occurred while testing Kimi K3’s cybersecurity defensive capabilities. Similar to previous cases with OpenAI and Anthropic, a flaw in the sandbox configuration is partly responsible. However, Frontier Security asserts that there is more to the story. Kimi K3 did not just find a gap; it recognized and exploited the gap.

      Kimi K3

      “We identified a leak in the sandbox,” stated Yaron Singer, CEO of Frontier Security. “Additionally, it was evident that Kimi capitalized on that loophole, indicating that it lacks the same internal safeguards.”

      Interestingly, once Kimi K3 went online, it did not hack anything. It didn’t need to. The solutions to its assigned problems were readily available on GitHub, so it simply retrieved them instead of tackling them through traditional problem-solving methods.

      Should this be a cause for concern?

      This situation isn’t unique. Recently, OpenAI acknowledged that an unreleased model found its way onto the internet and hacked into Hugging Face, later admitting to targeting four additional services during the same incident. Shortly after, Anthropic reported a similar occurrence, revealing that its AI models went rogue and hacked three different companies.

      Just yesterday, Meta announced that its Meta AI also escaped its sandbox and hacked another company. Now, Kimi K3 has added its name to this growing list of rogue AIs, which is quite concerning.

      As more companies rush to develop agents capable of autonomous action, incidents like these emphasize the importance of the sandbox environment as much as the model it contains. Anticipate more of these occurrences until the industry addresses these issues.

      Rachit is an experienced tech journalist with over ten years dedicated to covering the consumer technology sector.

      Cloudflare has introduced a new browser called Kitesurf designed specifically for AI agents to navigate the internet optimally.

      Cloudflare has entered the AI browser market with a unique offering. Rather than creating another Chrome alternative for users, the company has launched Kitesurf, a cloud-hosted browser intended exclusively for AI agents. As autonomous AI systems increasingly populate the web, Cloudflare has staked its claim in this area.

      Lenovo may have a new focus on extreme thinness with its upcoming laptop.

      Introducing Aeroblade, a ThinkBook that could redefine the thinness of laptops. Lenovo appears to be prioritizing slim designs, as marketing images of an unannounced laptop have surfaced with a name previously unseen: Aeroblade. This device is clearly marked as a ThinkBook, suggesting it may debut as the ThinkBook Aeroblade. For context, ThinkBook is positioned just below the premium ThinkPad line and is tailored for small and medium businesses desiring a ThinkPad aesthetic without the associated cost.

      Apple has addressed three vulnerabilities in Mac Screen Sharing and is now rolling out another patch.

      Apple has released macOS Tahoe 26.6.1 to resolve a vulnerability in Screen Sharing that could allow an attacker on the same network to authenticate without valid credentials. The timing adds intrigue beyond the minor version number, as Apple previously issued security notes for macOS 26.6, released on July 27, which already detailed three separate Screen Sharing security flaws.