Herding agents: Cybersecurity's new grand challenge

Welcome to Runtime! Today: As agents make their way deeper into the enterprise, security tools are trying to keep pace, Arm steps out from behind the curtain with its own server chips, and the latest funding rounds in enterprise tech.

Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime each week, and if you value independent enterprise tech journalism, click the button below and become a Runtime supporter today.

Rethinking security, again

The cybersecurity industry exists because tech workers have been making mistakes or invalid assumptions while writing software or configuring infrastructure for generations, an inevitable consequence of even the most well-trained and diligent engineers having a bad day or a brain fart. As enterprise tech shifts into a world where machines are doing a lot of that work, the problem could get worse before it gets better.

Security professionals descended en masse on San Francisco's Moscone Center this week for RSA 2026, the most corporate-friendly event on the cybersecurity calendar. There was a common theme among several new products and services unveiled during the first two days of the event: Security companies believe their customers are scrambling to keep up with new types of threats posed by AI chatbots and agents.

• Several new products, such as CrowdStrike's Falcon Data Security, were designed to address the age-old problem that CISOs face in trying to monitor activity on their networks.
• AI agents throw gasoline on this fire, moving large amounts of data at a rapid pace across multiple environments.
• Saviynt, which just raised $700 million in new funding, released a new product that helps companies update their identity security strategies for agents, which behave in new and unpredictable ways.
• Part of the problem is that agents can have multiple identities — such as the ID of the employee who built the agent as well as the access privileges temporarily delegated to the agent itself to complete its task — and a lot of identity security tools currently in use were designed to track one person's activity, said Brian Delahunty, vice president of AI for Google Cloud, in a recent interview with Runtime.

Security companies and data companies have been on a collision course for several years as it becomes clear that the only way to manage increasingly complex enterprise environments is to apply advanced data analysis to security alerts and vulnerability detection. That trend resurfaced at RSA 2026 through several different announcements.

• Databricks jumped into the security market for the first time with the release of Lakewatch, which brings several different types of data together in a single view as a new type of SIEM (Security Information and Event Management) product.
• Oracle added a new data access control system to its flagship database that shifts identity management decisions from applications to the database itself.
• And Elastic unveiled a new tool called Elastic Workflows that runs inside its Elastic Security service to allow customers to automate responses to security incidents based on data.

Agentic security tools still appear to be helping defenders more than attackers, but enterprises have been bracing for an onslaught of agent-driven attacks for several years, and at some point they'll arrive. Zenity's Michael Bargury gave a talk on Monday demonstrating how easy exploiting AI agents can be, according to The Register.

• Most successful cyberattacks over the last several years involved a great deal of social engineering, such as impersonating a help desk worker to gain access to systems or enticing employees to click on links that deliver malware.
• But attacks on "gullible" agents can be pulled off with clever prompts that allow attackers to take over systems without having to convince a real human to take action, Bargury told The Register.
• Companies that deploy agents at scale will need to be very clear and direct about what they allow those agents to do and where they are allowed to reach, and as we've seen many times over the years, there are probably going to be a lot of mistakes during the early rollout.

Did you have to call it that

Arm has been the company-behind-the-company for decades as a designer of high-performance chip cores that can be found in everything from the iPhone to server chips like AWS's Graviton and Nvidia's Vera. For years it was content to license those cores rather than compete with its customers with silicon of its own, but that changed Tuesday.

The Arm AGI CPU is a new server chip that was "designed to support massively parallel, high-performance agentic workloads in a densely populated rack deployment," the company said in a blog post. New Street Research's Pierre Ferragu told The New York Times that the announcement was "the most significant strategic pivot in the company’s history.”

Meta co-developed the chip with Arm and will deploy it in its data centers when it is released later this year. While several other prominent Arm customers —  including The Big Three hyperscalers — provided supporting quotes for the launch, check back in a year to see how many of those customers are actually using the chip.

Enterprise funding

XBOW raised $120 million in Series C funding for its "autonomous offensive security service, which is an AI-era take on penetration testing.

Oasis Security landed $120 million in Series B funding for its agentic identity-management service, which as discussed above is quickly becoming an issue for companies deploying agents at scale.

Dash0 scored $111 million in Series B funding for its observability technology, which was built around the OpenTelemetry project.

Gimlet Labs raised $80 million in Series A funding for its AI inference software, which helps companies run inference across multiple types of AI chips.

Andromeda landed $60 million in new funding for its compute marketplace, which connects buyers with sellers of excess GPU capacity.

Doss scored $55 million in Series B funding for its "operations cloud," which makes it easier to adapt older ERP systems around new projects and goals.

The Runtime roundup

AWS was forced to deal with new issues at its cloud region in Bahrain Monday amid the ongoing war in Iran, according to Reuters, although it wasn't clear if the data centers were hit by drones directly or if drones were causing problems in the area affecting its status.

Microsoft snatched up the 700 megawatts of computing capacity in Texas abandoned by OpenAI and Oracle earlier this month, according to Bloomberg.

Thanks for reading — see you Thursday!