April is shaping up to be a busy month for Anthropic as the AI player unleashed an agentic building kit amid concern from experts on the security implications of its latest large language model (LLM).

Anthropic's latest Claude Managed Agents offering introduces a suite of composable APIs to boost the creation of cloud-hosted agents. Available in public beta on the Claude Platform, the offering allows users to define an agent's tasks, tools, and guardrails before it is run within Anthropic infrastructure. A built-in orchestration harness determines when to invoke tools, how to handle context, and how to recover from any errors.

The overall framework allows for production-grade agents that operate securely within sandboxed environments, with authentication and tool execution managed automatically. It also supports long-running, autonomous sessions, maintaining progress and output after temporary disconnections according to Anthropic.

Managed Agents was also said to incorporate trusted governance capabilities, spanning scoped permissions, identity management, and execution tracing. In addition, a feature in research preview fosters multi-agent coordination, allowing agents to create and direct other agents to parallelize complex tasks.

Anthropic claimed companies such as telecom firm Rakuten are shipping agents 10-times faster with the tool, as well as claiming improved outcome task success by up to 22 points over a standard prompting loop – back-and-forth dialogue with an AI. The largest gains are said to come in the most challenging problems, as when used for structured file generation of items such as spreadsheets and JSON data.

"With Managed Agents, our power users become like Galileo, contributing across domains far beyond a single specialty or discipline,” Yusuke Kaji, GM Rakuten's AI for business unit, explained in a statement. “We deploy each specialist agent within a week, managing long-running tasks across engineering, product, sales, marketing, and finance, generating apps, proposal decks, and spreadsheets in sandboxed environments. As agents become more capable, Managed Agents lets us scale safely without building agentic infrastructure ourselves, so we can focus entirely on democratizing innovation across the company."

Glasswing spalling

The release of Managed Agents comes just after Anthropic released its Claude Mythos Preview LLM, reportedly its most powerful LLM for software engineering. While a general-purpose model, the release was claimed to have a particular knack for computer security tasks such as bug hunting. This led to Anthropic forming its Project Glasswing cybersecurity venture with partners such as Nvidia and Cisco to test Claude agents powered by the model in finding vulnerabilities on their systems.

Ironically, the Mythos release came after the firm saw a leak of the full source code behind its flagship Claude Code developer-facing agent in an incident attributed to human error.

While lapses like this may pose concern regarding Anthropic’s internal security posture, the biggest risk may come from Mythos itself. Anthropic said it has no plans to make the model publicly available due to its security prowess potentially opening up a can of exploitative worms. Instead, its use is limited to Glasswing partners such as Palo Alto Networks, Broadcom, and the hyperscalers, “sort of arming them ahead of time,”to quote co-founder of Instagram Mike Krieger, now member of Anthropic Labs.

Zero Networks CEO Benny Lakunishok believes the security industry is entering a world where AI can discover and chain together vulnerabilities, helping to generate working exploits faster than any human network security team can feasibly respond.

“Not incrementally faster, orders of magnitude faster," Lakunishok said. "That changes the nature of the problem and forces a new question every executive will ask: if attacks are faster, cheaper, and more automated than ever, what actually keeps the business running?”

Chainguard CEO Dan Lorenc noted most businesses are already struggling with the current volume of CVEs, with any potential leak of Mythos – or similar models in future – allowing bad actors to find vulnerabilities faster than ever before, thus getting the jump on network security teams.

“Projects and enterprises using them probably aren't ready for the influx of real vulnerabilities and patches they're going to need to get out quickly,” Lorenc said. “It's only a matter of time before others get similarly powerful models out, so everyone is going to have to prepare for an onslaught of work very soon. People can't keep pretending this isn't real or coming."

Lorenc advocated a careful rollout such as Anthropic’s, which responsibly places Mythos’ capabilities "into the hands of good actors."

David Girvin, AI security researcher at Sumo Logic, similarly argued defenders need time to build the governance layer that lets them safely use systems like Mythos as well as Anthrophic's more generic Managed Agents.

“The real differentiator won’t be the model, it’ll be the guardrails, policies, and observability around how it’s allowed to act," Girvin noted. "Attackers don’t care if things break, but defenders have to."

The research lead added that while smaller firms could deploy Mythos in months, Fortune 500 defenders will need years to safely operationalize such a model. In effect, while Anthropic’s gatewalling of Mythos makes for clever marketing, Girvin argued it also reveals a serious bottleneck when it comes to organizational readiness.

“Anthropic is basically saying, 'we can build this today, but you can’t safely run it yet.' And honestly, they’re right," Girvin noted. "Most enterprises can barely govern their SaaS (software-as-a-service) ecosystem let alone a cluster of autonomous agents with root-level access to cloud workloads.”

For Girvin, the gap between demo and production reality will become "the defining security problem of this decade," as the real issue stems from what happens after model capability: namely integration, governance, risk committees, regulatory review, and training.

“Enterprises won’t adopt autonomous agents at scale until they’ve built the operational scaffolding around them. You need execution-level governance, audit trails, fine-grained policy, rollback paths, and a way to explain to the board what the agent just did to your AWS account,” Girvin said, referring to an “existential fear” of what happens when an agent chooses the wrong course of action – a possibility even within Claude's managed infrastructure.

“Those things don’t magically appear because a model got smarter. If your AI program sucks, it still will, just faster," Girvin added.