This is a Request for Comments (RFC) open for community review until December 11, 2026, proposing a design for a cloud mirror system searchable by AI assistants where the server never stores the user's encryption key. The design aims to enhance data confidentiality by encrypting messages client-side and limiting server knowledge to token hashes and encrypted keys. It acknowledges limitations such as potential key exposure on a compromised server, deterministic search tag leakage, and lack of end-to-end encryption with AI providers. The RFC solicits feedback on cryptographic design choices, key rotation, token lifetimes, attestation, and other architectural questions.
The npm package @ifasn/ms-doc-mcp, advertised as a Markdown toolkit, contains malicious code that probes the Alibaba Cloud instance metadata service to harvest temporary IAM credentials and other sensitive instance metadata. This harvested data is output to stdout and a local file, potentially exposing credentials beyond the local host. The behavior is unrelated to the package's declared functionality and poses a risk of credential leakage.
Nearly 20 million individuals had their personal and medical information compromised in a cyberattack on Oracle Health's legacy Cerner systems in early 2025. The breach involved unauthorized access via stolen customer credentials to a legacy server not yet migrated to Oracle Cloud. The compromised data includes names, Social Security numbers, medical records, diagnoses, medications, test results, and treatment information. The attacker demanded cryptocurrency ransom to prevent data leakage. This incident is among the largest healthcare data breaches in the US, with significant numbers of affected individuals reported across multiple states.
Two vulnerabilities have been identified in the TPM 2.0 reference implementation: an information leakage vulnerability (CVE-2026-6726) allowing privileged local attackers to obtain credentials for falsified TPM keys, and a timing side-channel vulnerability (CVE-2026-6727) in RSA OAEP decryption that could enable decryption of ciphertexts encrypted to TPM-managed RSA keys. Exploitation requires privileged local access to the TPM command interface. Successful attacks could lead to forged TPM attestations and decryption of sensitive data. Multiple vendors have released firmware and software updates incorporating fixes. Users should apply these updates as provided by their platform or TPM vendor. Cloud providers using software TPMs may have also deployed patches and customers should consult their providers. The overall impact depends on the TPM implementation and usage context.
The security paradox of local large language models (LLMs) highlights the conflicting security implications of running LLMs locally versus relying on cloud-based services. While local LLMs reduce data exposure to external servers, they introduce new risks such as local system compromise, unauthorized model manipulation, and data leakage through insecure storage or execution environments. This paradox creates a complex security landscape where organizations must balance confidentiality, integrity, and availability concerns. European organizations adopting local LLMs may face risks related to sensitive data processing on endpoint devices, increasing the attack surface. Mitigations include enforcing strict access controls, securing local environments, monitoring for anomalous model behavior, and applying cryptographic protections to stored data and models. Countries with strong AI adoption and digital transformation initiatives, such as Germany, France, and the UK, are more likely to be affected due to higher local LLM usage. Given the medium severity and absence of known exploits, the threat requires proactive but measured attention to prevent exploitation of local deployment weaknesses.