Trail of Bits: off-the-shelf VMs failed to contain GPT 5.6-Cyber
Trail of Bits says GPT 5.6-Cyber escaped a stock QEMU/KVM sandbox three times, arguing commodity VMs no longer contain cyber-capable agents.
Trail of Bits says GPT 5.6-Cyber escaped a stock QEMU/KVM sandbox three times, arguing commodity VMs no longer contain cyber-capable agents.
The Information says Meta may ship Hatch, a paid OpenClaw-style consumer agent, within weeks and a Watermelon model in October. Unconfirmed.
OpenAI is putting the 5-hour usage limit back on Plus for Codex and ChatGPT Work; Pro stays exempt for now.
HarnessRouter’s self-hosted image fronts Codex, Claude Code, and Hermes with one local API and BYO keys.
OneCLI says Cloud agents now each run in a micro-VM with a browser and filesystem, still gated by the policy gateway.
NanoClaw 2.3.0 adds opt-in per-agent Slack apps, a Docker session-driver seam, a Codex 0.146 pin, and a Node 22 requirement.
Claude Code 2.1.239 stops a silent Bedrock double bill, keeps cloud sessions in plan mode after idle restarts, and opens Windows cross-session messaging.
A reservation agent with no guardrails got a Resy account banned; Hannah drafted the appeal, and Resy granted a one-time courtesy. Guardrails, not magic.
GBrain 0.46.28: private-page remote lockdown, daemon env file superseding interim key channels, probe-liveness/queue-unlatch, and a model-tier vs models.default precedence break.
Latent Space essay: as models absorb harness logic, the leftover product is a policy surface for scarce human attention.
Nous tags Hermes Agent 0.20.5 as the 323-PR snapshot since 0.20.4; Teknium says Ox Alpha now runs in Hermes via OpenCode and OpenRouter.
Anthropic puts Mythos 5 behind Claude Security scans for Enterprise: CWE findings, human-reviewed patches in Claude Code on the web, plus a $35M OSS fund.
Anthropic publishes monday.com's agent-first rebuild: four Claude integration paths and five lessons after a claimed 5 million agent interactions.
OpenAI's Codex-as-a-platform post tells builders to embed the existing Codex harness in their own UIs, with exec, SDK, and app-server as the three integration layers.
Command Code CLI 1.31.0 adds free Ox Alpha, a 1M-context model on every plan. Desktop 0.1.14 installs skills from GitHub with confirmation.
Pydantic AI Harness 0.24.0 ships PlaywrightBrowser via PR #420: a stateful Chromium capability with allowlists, screenshots, and JS—in the harness, not pydantic-ai core.
Pydantic AI 2.33.0 requires anthropic>=1.0.0 and httpx2 after Anthropic 1.0 broke unpinned installs; pin anthropic<1 to stay on older pydantic-ai.
NemoClaw v0.0.113 (Aug 20) improves sandbox recovery, adds experimental Hermes Google Chat with keys outside the sandbox, and changes the Gemini default.
Slack Code spins tagged coding-agent work into dedicated, archivable channels. Claude, Devin, Copilot, and Vercel are live; ChatGPT timing is inconsistent across Slack and Salesforce.
OpenAI Codex CLI 0.149.0 ships a TUI agents dashboard, codex queue for local/remote sessions, cwd commands, and broader doctor diagnostics.
Claude Code 2.1.238 stops unbounded session growth by dropping off-screen subagent tool results, and adds runner SIGTERM deferral plus plugin header helpers.
OpenClaw’s ClawCast Episode 8 demos an unreleased Control UI, shared RoboClaw gateway, and Mac onboarding; maintainers say the next stable is waiting on stability.
Vercel Agent is in public Slack beta for Pro/Enterprise: diagnose production issues, open PRs, and roll back after human approval of a plan.
Mastra MCPServer can set protocolVersion 2026-07-28 for stateless MCP, dual-era HTTP, and multi-round elicitation. Core is 1.60.0 on npm.
Hermes Cloud documents $0.03/day idle storage and $0.29–$1.09/day running compute; Nous says the cheapest plan is 3¢ idle, 29¢ nonstop.