KDCube
← AI Industry News

Edge Clouds Grab Agent Inference as an AI Agent Cracks Reactor Software

Akamai put a CDN-scale cloud behind agents with Akamai Inference and a WebAssembly Agent Sandbox, while Snowflake and GitHub shipped production controls for coding agents and portable, persistent Copilot sessions. Meanwhile Causal Security's autonomous agent Cipher found four zero-days — including a 9....

Lead signals

Highlights

  • Akamai put a CDN-scale cloud behind agents: Akamai Inference (12 open-weight models, 8B–2.8T params, OpenAI-compatible, serverless or dedicated) plus an Agent Sandbox that runs agent-written code in WebAssembly isolation with caps on compute, memory, outbound connections, and privileges (Techzine).
  • Snowflake gave its Cortex coding agent production knobs — a thinking_effort reasoning dial (minimal→max), appendable instructions.system, and stage_mounts to mount stages into the sandbox (Snowflake Docs).
  • GitHub made Copilot agent sessions persistent and portable — start on mobile, steer from VS Code, check from github.com — with an Agent Host that keeps work running after you close the window (GitHub).
  • Causal Security's autonomous agent Cipher found four zero-days in EPICS Base — the control toolkit behind ~90 accelerators and fusion reactors — in 3h49m, including a 9.8 RCE (Helix Threat Watch).

Key Signals

  1. Edge providers are now selling the agent runtime, not just the pipes — Akamai, Oct 28

    Akamai's beta pairs managed model serving (Nemotron 3 Ultra, gpt-oss-120b, Gemma 4) with an Agent Sandbox built on the Fermyon/Spin WASM framework, isolating generated code from the agent's own runtime. It extends the containment story from this week's microVM and Agent-Cloud entries into a cheaper, Wasm-based tier with default outbound limits — exactly the egress control operators have been asking for (Techzine).

  2. Data-platform agents get real orchestration controls — Snowflake, Oct 28

    The Cortex Agents agent:run API now accepts three public controls on requests using code_toolset_all: a reasoning-effort setting, extra system instructions that layer onto (not replace) the managed prompt, and stage mounts that expose versioned notebooks, Streamlit apps, and workspaces inside the sandbox. This is the difference between a demo coding agent and one teams can tune, govern, and point at governed data (Snowflake Docs).

  3. Offensive-security agents crossed into critical infrastructure — Causal Security, Oct 29

    Starting from an unfamiliar codebase with zero initial human intervention (one later authorization for deeper testing), Cipher traced network-controlled input to memory operations, reproduced four failures in a safe build, and surfaced CVE-2026-114595 (9.8, integer overflow → potential RCE) among three more. EPICS Base users should move to 7.0.11 (Helix Threat Watch, MayhemCode).

Why It Matters / What To Watch

  1. Inference and isolation are converging into one buy.
    • Compare Akamai Inference's serverless pay-per-use vs. dedicated pricing against your current model-serving bill before committing a runtime (Techzine).
    • Watch whether Wasm sandboxing (lighter than microVMs) becomes the default isolation tier for agent code execution (Techzine).
  2. "Coding agent" is becoming a configurable API surface, not a product.
    • If you run data-adjacent agents, test thinking_effort as a cost/latency lever and audit what stage_mounts exposes to the sandbox (Snowflake Docs).
    • Treat cross-device persistent sessions (GitHub Agent Host, cloud automations that self-improve) as a new audit boundary — background agents now run without the dev machine online (GitHub).
  3. Autonomous vuln discovery is now a sub-4-hour reality.
    • Assume adversaries can run Cipher-class agents against your open-source dependencies; prioritize patch cadence for input-parsing paths (Helix Threat Watch).
Quick links

Quick Links