
pieverse · 2026-11-13 04:49 UTC
Introducing Confidential Agents: Security Beyond the Wallet AI agents are becoming more capable, handling everything from everyday tasks to onchain interactions. But as agents take on more responsibility, they also accumulate sensitive information: conversations, files, credentials, and long-term memory. Protecting an agent's wallet is important. Protecting everything else it carries is the next step. Purr-Fect Claw already enables AI agents to interact with @BNBCHAIN. Now, we're extending our existing hardware-backed security infrastructure beyond the wallet and into the agent itself. Introducing Confidential Agents, a new Purr-Fect Claw option designed to protect an agent's data throughout its lifecycle, whether it's actively running or temporarily asleep. From Wallet Security to Agent Privacy Our approach to agent security began with the wallet. Purr-Fect Claw uses Trusted Execution Environments (TEEs) to protect wallet keys inside isolated hardware. Agents never hold those keys directly. Instead, transactions are processed through a separate signing environment, where user-defined spending limits and policies are enforced. But agents hold much more than wallet keys. They maintain conversation histories, store files, remember instructions, and manage credentials needed to interact with external services. Traditional encryption protects stored data and information in transit, but that data generally becomes readable again when software processes it. Those moments create additional points of exposure. Confidential Agents are designed to reduce that exposure by extending hardware-backed protection to the agent's runtime, memory, and stored data. How Confidential Agents Work Confidential Agents are built on Intel TDX and Google Confidential Space, combining hardware isolation, cryptographic verification, and per-agent encryption. A Trusted Execution Environment protects data while software is actively processing it. Before an agent's host can access protected keys, it must pass a process called hardware attestation, which verifies that it's running approved software in a trusted environment. Google Cloud Attestation verifies the underlying hardware evidence, while Pieverse's key service checks that the host meets the required security conditions. The platform must also authorize that host for the specific agent before its keys are released. The architecture combines three core protections: • Verified execution: Hosts must prove they're running approved software before receiving protected keys. • Per-agent encryption: Each agent has its own encryption keys for memory, messages, schedules, and sensitive data. • Isolated environments: Each agent operates in a separate software sandbox, with controls restricting access between agents and the underlying infrastructure. One important distinction: hardware isolation protects the host from cloud-operator access, while software sandboxes separate agents sharing that host. Each agent does not receive its own independent hardware enclave. Together, these protections are designed to limit unauthorized access to sensitive information without fundamentally changing how users interact with their agents. Protection That Continues While Agents Sleep An agent's security requirements don't end when a conversation does. Purr-Fect Claw's on-demand infrastructure allows agents to become idle and resume when needed. Confidential Agents extend encryption across that lifecycle. Each agent has its own root key, from which separate keys are derived for memory, incoming messages, schedules, and credentials. Working keys remain in memory and are cleared when an agent sleeps. Meanwhile, its memory archives and stored state remain encrypted. Messages received while the agent is inactive are also encrypted before storage. When the agent wakes, an authorized and verified host must receive the appropriate keys before the agent can recover its protected information. Certain sensitive credentials, such as supported skill API keys and bot tokens, can also be encrypted directly in the user's browser before reaching the agent. This allows Pieverse to store their encrypted form rather than readable copies. Routine support tools expose only content-free health information rather than private agent logs or conversation history. Confidential Agents are designed to support the same channels, skills, and tools as standard Purr-Fect Claw agents, with any exceptions disclosed before purchase. Understanding the Privacy Boundaries Confidential Agents represent an important step forward, but they don't make every connected service confidential. The first release focuses on protecting the agent environment and stored data. Some information still passes through other systems that need to process it. We refer to these points of exposure as the plaintext map: the places where information can still become readable during processing. • Message delivery: Connected messaging platforms can access messages they deliver. Pieverse's channel gateway also briefly processes incoming messages before encryption, while our relay holds live communication streams in memory during forwarding without storing or logging their contents. • AI model calls: Model providers can still read the prompts they process. Requests paid for by Pieverse pass through our current AI gateway, which retains an archive. Requests using a user's own provider key go directly to that provider, bypassing the Pieverse AI gateway. • Credentials: Not every credential is covered by browser-side encryption. Certain API and trading keys remain readable by the platform at launch. These boundaries will be disclosed before purchase. The security model also relies on Intel and Google's underlying infrastructure and does not eliminate every physical or insider threat. It is designed to reduce exposure through multiple layers of protection rather than claim absolute security. The distinction is important: Confidential Agents protect the agent itself, while broader end-to-end protection across the AI supply chain remains a separate area of development. Extending Protection Across the AI Supply Chain Confidential Agents are the next step in a broader effort to protect the systems connecting users, agents, and AI models. Looking ahead, we're working toward two additional infrastructure components: TEE AI Gateway would extend verified execution environments to message handling, model-call routing, and provider credentials, reducing the number of systems that can access readable information. TEE AI Inference would extend protection further by running supported AI models on confidential-computing hardware, protecting data during model processing. We also plan to publish signed software release records and a public attestation verifier to make these security controls more independently auditable. These are future developments, not features of the initial Confidential Agents release. No timelines are being announced for these components. Rollout and What's Next Confidential Agents are currently preparing for release. The underlying key-management infrastructure and certificate authority have already been tested on real Intel TDX hardware. Before users are onboarded, we're completing additional host enrollment, end-to-end validation, scalability and recovery testing, and an independent security review. The initial rollout will be limited to invited users, with broader availability dependent on further development and validation. Our approach remains the same: build security into each layer, verify what can be verified, and be transparent about what remains outside those protections. Confidential Agents are the next step in extending that approach across Purr-Fect Claw, and they're coming soon. Rollout starts next Monday.
Loading the story on Tree News…