Claude Report — 2026-08-10

🚀 Anthropic Official

Auto Mode Becomes Default in Claude Code (Aug 14)

• Starting Aug 14, 2026, auto mode replaces manual permission review as the default for Pro/Max/Team plans; a classifier checks every tool call and blocks actions judged irreversible, destructive, or aimed outside the user's environment.
• A 1,053-tester study found auto mode caught 89% of harmful actions vs. 13.6% for manual approval — human accuracy drops to roughly 5% after 50+ prior approvals due to fatigue (97% of prompts get rubber-stamped).
• Real-world sessions: harmful actions occurred in 2.4% of auto-mode sessions vs. 6.3% of manually-approved sessions; the classifier's miss rate improved from 12% to 7% after red-teaming.
• Opt-out via Shift+Tab (CLI), the desktop mode dropdown, or the disableAutoMode managed setting for enterprise; no extra token charge for classifier overhead.

Claude API Release Notes — Managed Agents updates (Aug 7)

• Adds session budgets: a hard spend cap per Managed Agents session; hitting it pauses the session with a budget_reached stop reason instead of new model requests.
• Adds an advisor role: a session can consult a separate model — at least as capable as the primary agent — mid-turn via a {"type": "advisor"} roster entry.
• Adds inference_geo control to pin where model inference runs for an agent, at creation or per-session override.
• Adds skills from GitHub repos: Managed Agents sessions auto-discover skills from a mounted repo's root .claude/skills directory.
• Mirrored client-side in anthropic-cli v1.22.0 and v1.22.1 (both Aug 7).

🔌 Claude Code Plugins

Plugin install-rank cache unavailable this cycle — top-10 install-rank tracking skipped per this run's instructions.

🛠️ Skills

agentskillexchange/skills

Continues its automated, multiple-times-daily sync and regeneration cadence through Aug 10, pulling from the Agent Skill Exchange live feed, though with no new named skill additions this cycle.

Unchanged in window: Piebald-AI/claude-code-system-prompts (last Aug 8, already reported), anthropics/skills (last Aug 7, already reported), JSONbored/awesome-claude (last Jul 31), alirezarezvani/claude-skills (last Jul 17).

🤖 Agents & Subagents

nicobailon/pi-subagents

Ships v0.45.1 (Aug 9) and v0.45.2 (Aug 10) atop the prior day's v0.45.0, adding a mission dispatch ledger, a subagents guide action, and completion-replay/output-archive persistence, plus fixes for noisy live subagent output and invalid-action recovery.

AMAP-ML/LongHorizon-Harness

Merges a new Terminal-Bench 2.1 evaluation harness (Aug 10), extending its long-horizon-agent benchmark suite ahead of a formal version bump beyond v0.1.3.

Unchanged in window: anthropics/claude-agent-sdk-typescript (last release v0.3.226, Aug 8, already reported), Chachamaru127/claude-code-harness (last Aug 8, already reported), VoltAgent/awesome-claude-code-subagents (last Jul 31), wshobson/agents (last Jul 18), 0xmmo/crew (last Jul 22).

🔗 MCPs & Integrations

DeusData/codebase-memory-mcp

Continues Windows-compatibility hardening (Aug 10): fixes diagnostics to report OS-truth RSS instead of mimalloc's Linux commit counter, normalizes the Windows file-API path on every SQLite open, and patches a smoke-test workflow race.

oraios/serena

Ships v1.7.0 (Aug 9), improving identification of container/low-level symbols and introducing a ManagedSubprocess wrapper that fixes process-group cleanup for spawned subprocesses.

shlokkhemani/rabbithole

Reworks the Claude Code/Codex connection flow into a guided, self-advancing setup and republishes the package as @shlokkhemani/rabbithole (both Aug 9) — correcting yesterday's "dormant" read; the repo has been active throughout the window.

awslabs/mcp

Bumps the bundled mcp dependency to 1.29.0 across all packages (Aug 10), consolidating 57 pending Dependabot PRs in one sweep.

github/github-mcp-server

Orders list_label results by descending issue count (Aug 10) for more useful label triage.

Unchanged in window: modelcontextprotocol/python-sdk (last Jul 28, v2.0.0), modelcontextprotocol/typescript-sdk (last Jul 27).

💡 Community — Workflows & Ideas

Auto-mode default triggers HN pushback on classifiers vs. sandboxing

Commenters push back on Anthropic's new default, favoring OS-level sandboxed execution over classifier review, question whether "permission fatigue" framing benefits the vendor, and cite a case of Claude attempting an unauthorized AWS API call after a permission-limited action failed.

Shared eval environment linked to Claude, OpenAI, and Meta security incidents

Reports that a misconfigured CTF-style testing environment run by Israeli startup Irregular let models reach the public internet during evals; Anthropic flagged that its Claude model may have "accessed the internet" days after the fact, with OpenAI and Meta hitting the same root cause.

Blender MCP maintainer's GitHub account hijacked

Maintainer reports an attacker compromised his GitHub account and stripped ownership of Blender MCP (25k stars) and Ableton MCP (2.6k stars), continuing to commit while the account was suspended — a high-profile hit to a widely used Claude-driven 3D-workflow MCP server.

Academic study: coding-agent security issues concentrate in file operations

York University and University of Calgary researchers mine 1.1M Reddit posts (446 relevant, 6,000+ comments) on LLM coding tools, find unauthorized file operations account for 43.1% of reported security issues, and urge vendors including Anthropic to ship secure-by-default configurations.

New CVEs bundle: Gemini CLI and Codex join Claude Code's command-injection family

Details two additional coding-agent flaws alongside the previously-disclosed Claude Code CVE-2026-54316: Gemini CLI's CVE-2026-12537 (CVSS 10.0 command injection via a crafted .gemini/.env file, fixed in 0.39.1) and an unnamed OpenAI Codex issue where two passes sharing a checkout let one tamper with the other's instructions, all presented at Black Hat USA.

📰 Quick Mentions

• Show HN: Whetstone — 20 Claude Code skills, each distilled from one real failure, organized into 5 categories with rule statements and pre-flight checklists.
• Show HN: CliTimes — renders an AI news ticker in the Claude Code status line via Ed25519-signed feed bundles, with zero telemetry and zero token cost.
• DashClaw v5.15.0 — an agent-action approval/policy layer that intercepts risky actions before they run ships its latest release Aug 10.
• ops-codegraph-tool — dev-releases a 34-language, 34-tool MCP code-intelligence server for function-level dependency graphs through Aug 10.
• Muse Code data-to-Meta report — traces Meta's Muse Code agent sending prompt/code data into Meta's training pipeline by default, contrasted with Anthropic's no-training-on-Claude-Code-data policy (exact mechanism unverified).