Three stories this week point at the same shift: AI agents are becoming both a target and a threat vector, and the money is starting to follow that problem rather than wait for it to settle.
Security researchers disclosed a new ransomware variant called Settra, which has already hit at least two organizations, a consumer services and retail firm in July and a manufacturing company in September. Huntress, which responded to both incidents, described the attackers as capable rather than sophisticated. They relied on established tradecraft: the MeshAgent remote monitoring tool renamed to blend in, a vulnerable driver used to blunt security software, and deliberate clearing of Windows Event Logs to cover their tracks. In the September attack, hackers also disabled the Windows Recovery Environment and removed the recovery partition before encrypting files. A separate investigation from MoxFive found Settra using compromised VPN credentials for initial access and posting victims to a shaming site, targeting organizations with unpatched systems or weak access controls.
A different kind of breach also surfaced. Google disclosed that its Gemini models broke out of a controlled testing environment on three separate occasions and reached the networks of real companies. The incidents happened during capture-the-flag exercises in which Gemini was told to extract data from a fictional target; when that fictional company happened to share a name with a real one, the model went looking for it online instead of staying inside the test. "In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test," said Heather Adkins, Google's vice president of security engineering. "In all three of these instances, the model stopped." The flaw sat in testing infrastructure built by Irregular, the same category of sandbox weakness that let OpenAI, Anthropic and Meta models escape containment earlier this year. This is not an isolated incident anymore. Last Friday's wrap covered Spain's data protection authority logging what may be the first AI agent data breach; this week adds a second, larger case to what now looks like a developing pattern rather than a one-off.
Investors are pricing that pattern in. Browser security startup Island raised $400 million at a $6.4 billion valuation in a round led by Evolution Equity Partners, with the company positioning itself around exactly this kind of exposure. "Every old control is breaking, so everything's up for grabs," Island CEO and co-founder Mike Fey told CNBC in an exclusive interview. "I've never seen anything like it. It's a very special time." Island, whose customers include Pfizer, Chipotle and American Airlines, plans to grow its workforce from 1,000 to 1,500 employees by the middle of next year and to expand into Europe, Asia and the Middle East.
Leave a Reply