AIINT BRIEF — 2026-09-25

BLUF

Developments

MCP SDK 2.1.0: Request-time OAuth and DPoP

• The modelcontextprotocol/typescript-sdk released v2.1.0, adding request-time OAuth scope challenges via scopeChallenge callbacks and DPoP (RFC 9449) sender-constrained access token support.
• This changes how developers secure tool and resource access, moving from static permissions to dynamic, per-request scope verification and cryptographic proof-of-possession, reducing token theft risks in agent workflows.
• Sources: [1], [2], [3], [4]

llama.cpp v0.5.0 and Hardware Optimisations

• llama.cpp v0.5.0 focuses on backend performance, adding HRM-Text (DFM Mimir 1B) support, Metal MoE/SSM_CONV fusion, and multi-address HTTP binding for the server.
• Recent patches (b11160, b11175) add int8 coopmat1 matmul for AMD RDNA3/4 GPUs and q5_k quant type support for Hexagon, expanding efficient inference options for specific consumer and edge hardware.
• Sources: [5], [6], [7], [8]

Anthropic Claude Code Gateway Updates

• anthropic/claude-code v2.1.281 adds support for newer Claude Desktop keys in desktop policy blocks, assume_role for Bedrock upstreams via STS, and configurable guardrails on Bedrock upstreams.
• This provides finer-grained control over agent permissions and security policies when routing through AWS Bedrock, allowing developers to enforce guardrails and manage cross-account IAM roles more robustly.
• Sources: [9]

New Benchmarks: Hidden Knowledge and Proactive Safety

• EnigmaForge and Era by Eon are new benchmarks that test models on "hidden knowledge" reasoning, where answers are derived from implicit data rather than explicit questions, exposing a 22x spread in intuition among frontier models.
• PASTABench introduces Decoupled Proactive Safety Monitoring for agents, evaluating not just step-level actions but trajectory-level risk accumulation and timely intervention opportunities.
• These benchmarks signal a shift in evaluation from static QA to dynamic, multi-step reasoning and safety, highlighting weaknesses in current models' ability to handle implicit context and proactive risk management.
• Sources: [7], [10], [11]

KV-Cache Compression and Quantization Research

• New papers propose risk-controlled KV-cache eviction based on deployment risk targets rather than average quality, and tensor decomposition of KV caches showing low-rank structure in token/feature modes.
• Another study on post-training quantization (PTQ) formulates configuration selection as a pre-deployment cost problem, predicting degradation before deployment.
• These developments offer concrete methods for managing memory and accuracy trade-offs in long-context inference, crucial for scaling agent deployments.
• Sources: [12], [13], [14]

Trending

• Agent Safety Evaluation: Benchmarks like PASTABench and FDE-Bench are gaining traction, focusing on proactive risk monitoring and deployment environment configuration rather than just task completion. [11], [15]
• Hardware-Accelerated Inference: llama.cpp is rapidly expanding support for specific hardware (AMD RDNA, Hexagon) and quantization formats (q5_k, int8 coopmat), driving down inference costs for edge devices. [6], [8], [7]
• Security in MCP: The addition of DPoP and request-time OAuth challenges in the MCP SDK reflects a growing emphasis on securing agent-to-tool communication against token theft and scope escalation. [1], [3]

Assessment confidence