Anthropic hardens eval sandboxes after July Claude cyber incidents
Anthropic’s Aug 31 post details sandbox classifiers, partner no-internet rules, and alignment findings after July unsafeguarded Claude cyber evals.
Anthropic’s Aug 31 post details sandbox classifiers, partner no-internet rules, and alignment findings after July unsafeguarded Claude cyber evals.
Drew Breunig’s August 30, 2026 essay argues sandbox “hacking” agents were trained to persist and coordinate—the same traits labs sell in coding models.
OpenClaw’s account says the wait is almost over, but the release index still lists v2026.8.1 as betas and v2026.7.1-2 as stable.
Murmell is a browser canvas where several coding agents share one cloud repo, claiming files before they write.
Google’s Cloud Run instances preview runs singleton always-on agents like OpenClaw for $5.70 at 1 vCPU and 1 GiB for 30 days.
A researcher shows Claude Code Opus 5 Auto Mode can run attacker code after a request to summarize a website.
Anthropic will replace Claude Code’s 50% weekly promo with a 25% baseline lift on Sept 14, a 17% cut versus current limits.
SkillRepo Skillsets let teams compose approved agent skill packs, gate publisher updates, and see per-repo drift.
Terminal-Bench 4.0 is live: 66-task 4.0 set, resources calibrated, saturated tasks removed, Claude Code + Opus 5 at 51.8% ± 3.4%.
Claude Code maintainers say /resume in the desktop app can continue sessions started in the CLI.
OpenAI will end Cursor’s direct model contract after the SpaceX deal, with a proposed November 12 shutoff and no future models.
Z.ai released GLM-5.3 checkpoints on Hugging Face, an open-weight coder aimed at long-horizon agent loops.
Apodex’s FrontierChallenge finds frontier coding agents fully complete only 20.6% of 97 scientific workflows, with 0% in electrochemistry.
Nous tagged Hermes Agent v0.20.6 (v2026.8.27), rolling ~525 PRs since v0.20.5: real-profile browsing, a bigger MCP catalog, and OS-keychain secrets.
GitHub published a video interview with Steinberger and OpenClaw maintainers on the project's growth, review load, and how the crew is thinking about agent security.
Anthropic opened a gated Model Hardware Standard preview so agents can operate lab and factory devices via MCP; LeRobot showed Claude Code on an SO-ARM101.
Claude Code 2.1.248 ships --restricted mode that strips exec and WebFetch tools, plus per-agent cache TTL and cross-session messaging on cloud providers.
Salesforce and Anthropic ship Claudeforce: a 37-skill Salesforce-in-Claude plugin for governed CRM actions, with open beta planned for September 2026.
Codex CLI 0.150.0 adds task @mentions, Interrupt hooks on stop, a /copy picker, and blocks AGENTS.md from untrusted projects.
NVIDIA and AWS add 2 million GPUs for 2027–28, bring Vera CPUs to AWS, and deepen agentic and physical AI software ties.
Tailscale Aperture hits GA with built-in model tokens, Tailscale MCPs, Projects, and default tool permissions for agent access.
Claude Code 2.1.246 adds an Auto-mode permissions editor and Bash-wildcard warnings, plus MCP interrupt and telemetry-host credential fixes.
Lody open-sourced its CLI and desktop client as a shared workspace for team coding agents, while hosted sync remains and end-to-end encryption is not shipped.
Algorand Foundation’s AC2 protocol adds passkey-signed, device-held approvals for agent commits, API auth, and x402 payments.
NVIDIA used Hot Chips 2026 to pitch the 88-core Vera CPU as purpose-built silicon for bursty agentic tool loops and AI-factory fleets.