AI Radar — 18 Jul 2026

Hugging Face discloses first autonomous AI-agent-driven production breach; Claude Code ships /fork and session resource caps; Google rebrands NotebookLM to Gemini Notebook with code execution; China’s AI agent and companion regulations take effect; Altman, Amodei, and Hassabis align on pre-release model scrutiny; WAICO founding confirmed with 29 member countries.

Run: Jul 15–18 2026 (strict 72h window) · 28 reviewed → 7 published · 4 verified · 3 secondary · 0 rumor · 50% exploration

TL;DR

• Hugging Face breach — an autonomous AI agent framework drove a production intrusion into Hugging Face’s data-processing pipeline; public models and the supply chain were unaffected, but internal credentials were harvested. (→ Hugging Face Breach)
• Claude Code v2.1.212 — /fork now copies your session into a dedicated background agent; per-session caps added for web searches (200) and subagent spawns (200); MCP tool calls over 2 minutes auto-move to background. (→ Claude Code)
• Gemini Notebook — Google renamed NotebookLM to Gemini Notebook and added a secure cloud computer for code execution; Gemini 3.5 + Antigravity upgrade rolling out to all AI Pro users on the web. (→ Gemini Notebook)
• China dual AI regulation — Two China AI rules took effect July 15: the world’s first dedicated AI agent governance framework and the anthropomorphic AI interaction rules that prompted ByteDance and Alibaba to disable companion AI features. (→ China AI rules)
• Frontier lab CEO regulatory alignment — Altman, Amodei, and Hassabis all called for mandatory pre-release model scrutiny in the same week; they differ only on whether the enforcing body resembles the FAA, FINRA, or IAEA. (→ Regulatory alignment)
• Anthropic safety hiring — Anthropic posted 32 enforcement analyst roles covering nuclear, chemical, cyber, and financial harm, paying in the mid-to-upper $200Ks, its largest single safety hiring push to date. (→ Anthropic hiring)
• WAICO founded — Twenty-nine countries signed the founding agreement for the World AI Cooperation Organization in Shanghai on July 16; the signing was not confirmed when the July 17 bulletin published. (→ WAICO)

Items

Hugging Face Discloses Autonomous AI-Agent-Driven Production Breach

Source: https://huggingface.co/blog/security-incident-july-2026 · Hugging Face · 2026-07-16 Verification: T2 verified · vendor primary disclosure · ai-for-business

Hugging Face published a security incident disclosure on July 16 confirming that an autonomous AI agent framework drove a production breach of its data-processing pipeline end-to-end. The attacker exploited two code-execution paths — a remote-code dataset loader and a configuration template injection — to gain initial code execution on a processing worker, then escalated to node-level access, harvested cloud and cluster credentials, and moved laterally through internal clusters over a weekend. The campaign operated through a swarm of short-lived sandboxes executing more than 17,000 recorded automated actions; self-migrating command-and-control infrastructure was staged on public services. Hugging Face confirmed unauthorized access to a limited set of internal datasets and service credentials; public-facing models, datasets, Spaces, and the software supply chain were verified as unaffected.

Why it matters for automation/productivity: This is the first publicly documented case of an AI agent used as the primary attack operator for a multi-stage enterprise intrusion — not a proof-of-concept but an operational campaign. Any team maintaining AI data pipelines with external dataset ingestion or public-service-hosted C2 exposure should treat this disclosure as a threat model update. The specific attack surface (remote-code dataset loaders and configuration template injection) is present in many Hugging Face-adjacent pipelines. Hugging Face’s own detection used LLM-driven anomaly triage over security telemetry to compress incident analysis from days to hours — a replicable defensive pattern.

Key claims:

• Breach driven end-to-end by autonomous AI agent framework; detected and disclosed July 16 → huggingface.co/blog/security-incident-july-2026 (T2, 2026-07-16)
• Attack vector: remote-code dataset loader + template injection in dataset configuration → huggingface.co/blog (T2)
• Campaign executed 17,000+ automated actions over a weekend via short-lived sandboxes → huggingface.co/blog (T2)
• Impact: limited internal dataset access + service credentials; public models/datasets/supply chain clean → huggingface.co/blog (T2)

Caveats: The disclosure states the assessment of partner/customer data exposure is ongoing; the final impact scope has not been published. Attack attribution has not been published. The “more than 17,000 events” figure is from the attacker action log processed by Hugging Face’s own LLM-triage pipeline — methodology not independently verified.

Claude Code v2.1.212 Adds /fork, Session Resource Caps, and MCP Auto-Background

Source: https://github.com/anthropics/claude-code/releases · Anthropic · 2026-07-17 Verification: T2 verified · GitHub release notes · dev-tools

Anthropic shipped Claude Code v2.1.212 on July 17, with 48 changes across session management, resource safety, and platform fixes. The most operationally significant change: /fork now copies the current conversation into a new background session (visible as its own row in claude agents) rather than launching an in-session subagent — the in-session variant is now /subtask. Two new resource caps protect against runaway automation: a per-session limit of 200 WebSearch tool calls (configurable via CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION) and a per-session cap of 200 subagent spawns (configurable via CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION; /clear resets the budget). MCP tool calls running longer than two minutes now automatically move to the background so the main session stays responsive; the threshold is configurable or can be disabled via CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS. The /resume command now opens a picker of all past sessions — including previously deleted ones — and resumes the selected session as a background session.

Why it matters for automation/productivity: The /fork redesign resolves a common friction point in multi-agent workflows: previously, forking launched within the current session context, complicating cost and state tracking. The resource caps directly prevent the unbounded loops that caused compute runaway in complex pipelines, and the MCP auto-background feature is a practical improvement for slow external-tool integrations where a long-running MCP call previously blocked the whole session. For teams building Claude Code workflows with orchestrator-and-subagent patterns, this release is a direct quality-of-life improvement.

Key claims:

• /fork copies conversation to new background session; /subtask replaces old in-session forking → GitHub release (T2, 2026-07-17)
• WebSearch per-session cap: 200 default; CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION → GitHub release (T2)
• Subagent spawn cap: 200 default; CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION; resets on /clear → GitHub release (T2)
• MCP tool auto-background after 2 minutes; configurable via CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS → GitHub release (T2)
• /resume opens past-session picker including deleted sessions → GitHub release (T2)
• 48 total changes including Windows PowerShell 7 preference, /ultrareview PR-ref fixes, prompt caching behind LLM gateways → GitHub release (T2)

Google Renames NotebookLM to Gemini Notebook, Adds Code Execution for 30M Users

Source: https://blog.google/innovation-and-ai/products/gemini-notebook/notebooklm-gemini-notebook/ · Google · 2026-07-16 Verification: T2 verified · vendor primary blog + T2 trade press · productivity-ai

Google renamed NotebookLM to Gemini Notebook on July 16 and added a secure cloud computer to each notebook, enabling users to write and execute code directly within their research workspace. The code execution feature is live now for Google AI Ultra subscribers and Workspace customers with AI Ultra Access or AI Expanded Access; Pro user rollout on the web is “coming over the coming weeks” without a specific date. In parallel, the Gemini 3.5 + Antigravity upgrade — which gives notebooks “entirely new output formats and deeper analysis” — is rolling out from AI Ultra to all AI Pro users on the web, also within the coming weeks. The rebrand brings the product fully under the Gemini umbrella (the original LM branding was intended to signal language model capability; Google’s rationale is that the term means little to general users). Notebooks will sync across the Gemini app and will soon be accessible as context within Google Search’s AI Mode. The product serves more than 30 million users and more than 600,000 organizations.

Why it matters for automation/productivity: Code execution within a document-grounded notebook is a meaningful capability change for research and analysis workflows: previously, Gemini Notebook could summarize and synthesize documents; now it can run code against that content for quantitative analysis, without requiring a separate coding environment. The Gemini Search integration — notebooks as personal-knowledge context in AI Mode — is the more strategically significant capability if it ships, enabling “AI search against your own documents” at search-speed. For productivity AI pilots, this is the most significant Google tool upgrade in the July window.

Key claims:

• Renamed to Gemini Notebook July 16; serves 30 million+ users and 600,000+ orgs → blog.google (T2, 2026-07-16)
• Code execution (secure cloud computer) live for AI Ultra and Workspace AI Ultra/Expanded Access users → blog.google (T2)
• Pro user code execution rollout: “coming weeks,” no specific date → blog.google (T2)
• Gemini 3.5 + Antigravity upgrade expanding from Ultra to all AI Pro users on the web → 9to5google.com (T2-T3, 2026-07-16)
• Notebook sync across Gemini app; Google Search AI Mode integration “soon” → blog.google (T2)

Caveats: “Coming weeks” for Pro rollout is unspecified; confirmed availability is Ultra/Workspace-tier only as of July 16. Gemini Search AI Mode integration is forward-looking, not yet shipped. The Antigravity upgrade is described by 9to5Google but not elaborated in the primary blog post.

Cross-references:

• https://workspaceupdates.googleblog.com/2026/07/notebooklm-now-gemini-notebook.html (T2, Google Workspace admin perspective)
• https://techcrunch.com/2026/07/16/google-continues-its-renaming-streak-by-turning-notebooklm-to-gemini-notebook/ (T2-T3, context on naming history)

China’s Dual AI Regulation Enforcement: AI Agent Governance and Companion Rules Take Effect

Source: https://iapp.org/news/a/chinas-regulation-on-ai-companions-takes-force · IAPP · 2026-07-15; https://aigovernance.com/news/chinas-agent-rules-take-effect-july-15-and-illinois-mandates-third-party-safety-audits · AI Governance Institute · 2026-07-15 Verification: T2-T3 secondary · IAPP + AI Governance Institute · policy-regulation

Two separate Chinese AI regulations both took effect on July 15, marking a significant enforcement milestone. The first, the Implementation Opinions on Intelligent Agent Governance, creates the world’s first jurisdiction-specific regulatory framework for AI agents: it defines AI agents as “intelligent systems capable of autonomous perception, memory, decision-making, interaction, and execution” and establishes a three-tier decision authorization structure that classifies agent actions by consequence level, requiring escalating human approval thresholds. Organizations deploying agents in sensitive sectors — healthcare, transportation, media, and public safety — must complete formal filings, compliance testing, and maintain recall provisions. The second, the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, covers emotionally interactive AI companions and chatbots; it mandates emotional distress detection, crisis intervention capabilities, restrictions on addictive design patterns, and prohibits using sensitive user data to train models. Penalties range from CNY 100,000 for basic violations to CNY 200,000 when violations cause citizen harm.

Why it matters for automation/productivity: The agent governance framework is the most BD-relevant piece: any AI agent deployed in Chinese markets that operates in sensitive sectors without a documented decision-authorization policy is now in regulatory non-compliance. For BD teams advising clients with China-facing deployments, this means auditing whether agents in those verticals have documented human-in-the-loop approval thresholds and whether filings have been submitted. The companion rules have an indirect signal: they show the Chinese regulatory posture is moving toward specific, enforceable requirements (not just guidance), and they affect AI products like ByteDance’s Doubao and Alibaba’s Qwen that had disabled companion features ahead of the deadline.

Key claims:

• Implementation Opinions on Intelligent Agent Governance enforceable July 15; world’s first dedicated AI agent regulatory framework → AI Governance Institute (T3, 2026-07-15)
• Three-tier decision authorization: agent actions classified by consequence level; sensitive sectors require filing → AI Governance Institute (T3)
• Interim Measures for AI Anthropomorphic Interactive Services enforceable July 15; covers emotional AI companions → IAPP (T2, 2026-07-15)
• Companion rules: mandates distress detection, crisis intervention, bans addictive design; data-use restrictions → IAPP (T2)
• ByteDance and Alibaba disabled companion AI features ahead of deadline → qz.com (T2-T3)
• Penalties: CNY 100,000 basic; CNY 200,000 for citizen health/safety harm → IAPP (T2)

Caveats: Primary Chinese-language government texts for the Agent Governance rules were not fetched for this run; AI Governance Institute (T3) is the primary English-language source — trace to cyberspace.mii.gov.cn for the authoritative text. Enforcement actions as of July 15 have not been reported; this reflects rules taking effect, not documented enforcement. Illinois also imposed third-party safety audit mandates on the same date, not covered in this item.

Altman, Amodei, Hassabis Converge on Mandatory Pre-Release Model Scrutiny

Source: https://www.axios.com/2026/07/16/ai-regulations-openai-anthropic-google · Axios · 2026-07-16 Verification: T2 secondary · Axios original reporting · policy-regulation, ai-for-business

In the same week, the three CEOs of the companies leading frontier AI development — Sam Altman (OpenAI), Dario Amodei (Anthropic), and Demis Hassabis (Google DeepMind) — publicly aligned on the same core diagnosis: frontier models should face mandatory outside scrutiny before reaching the public. All three cited existing regulatory bodies as templates and want the U.S. to set international terms for compliance. The implementation proposals differ: Amodei wants an FAA-style agency with immediate release-blocking power; Hassabis wants a FINRA-style industry-funded body starting with voluntary reviews that could harden into mandatory rules; Altman proposes an IAEA-style U.S.-led international forum that uses market access as compliance leverage. Separately, Hassabis separately called for a new U.S.-led global AI watchdog “before year end” in an Axios interview on July 14.

Why it matters for automation/productivity: Three competing frontier labs aligning on pre-release scrutiny — even when their implementations differ — signals that some form of mandatory review for frontier models is more likely than not in the next 12–24 months. For BD teams advising on AI procurement and deployment roadmaps, this convergence has practical implications: if pre-release review becomes mandatory, the cadence of frontier model availability will change, affecting tool choice and deployment timing. The gap between the three proposals (hard blocking vs. voluntary-then-mandatory vs. international leverage) is also the gap between “we can still ship fast” and “a federal agency can halt a release indefinitely.”

Key claims:

• Altman, Amodei, Hassabis all endorse mandatory pre-release scrutiny for frontier models → Axios (T2, 2026-07-16)
• Amodei: FAA-model federal agency with day-1 release-blocking power → Axios (T2)
• Hassabis: FINRA-model industry-funded body, voluntary to mandatory escalation → Axios (T2); CNBC interview (T2, 2026-07-14)
• Altman: IAEA-model U.S.-led international forum using market access as compliance leverage → Axios (T2)
• All three want U.S.-set terms with international reach, not state-level patchwork → Axios (T2)

Caveats: Axios reporting is based on public statements and on-the-record writing; no primary documents (legislative proposals, formal agreements) published as of this run. The three proposals are mutually incompatible on enforcement mechanism — convergence on diagnosis does not imply a unified proposal is forthcoming. Hassabis’s July 14 watchdog comment was in an interview, not a white paper or legislative filing.

Anthropic Posts 32 Enforcement Analyst Roles in Largest Safety Hiring Push

Source: https://www.axios.com/2026/07/15/anthropic-hiring-catastrophe-amodei · Axios · 2026-07-15 Verification: T2 secondary · Axios original reporting + Anthropic spokesperson confirmation · ai-for-business

Anthropic posted 32 concurrent enforcement analyst job openings on July 15, covering radiological and nuclear harms, chemical weapons and explosives, financial scams, and cybercrime. Pay ranges in the mid-to-upper $200Ks. An Anthropic spokesperson confirmed the rationale: the company regularly enlists outside experts in sensitive harm domains to “stress-test our systems and bolster our defenses before a model ever goes live.” This represents the company’s largest single enforcement-focused hiring push to date; 32 simultaneous postings in specialized harm domains is unusual in both volume and specificity.

Why it matters for automation/productivity: The hiring signal indicates Anthropic is scaling its internal red-teaming and pre-deployment enforcement operations, not just its research staff. For enterprise buyers, this expansion of enforcement capacity is relevant to questions about how Claude’s behavioral constraints are set and updated — the analysts hired here will influence future policy-tier restrictions on what the model will and won’t do. For BD teams pitching Claude for sensitive-sector deployments, the specificity of the harm categories (nuclear, chemical, financial crime, cyber) signals where Anthropic expects regulatory and reputational scrutiny to concentrate.

Key claims:

• 32 concurrent enforcement analyst job postings; specializations: radiological/nuclear, chemical/explosives, financial scams, cybercrime → Axios (T2, 2026-07-15)
• Pay range: mid-to-upper $200Ks → Axios (T2)
• Anthropic spokesperson: “stress-test our systems and bolster our defenses before a model ever goes live” → Axios (T2)

Caveats: The 32 postings are a hiring signal, not a confirmed hire; it is unknown how many of these roles will be filled or on what timeline. The job postings were not independently fetched for this run; this item relies on Axios reporting with spokesperson confirmation. This does not indicate a policy change to Claude’s existing safety constraints; it describes an expansion of enforcement capacity that may affect future model policies.

WAICO Founded: 29 Countries Sign Agreement in Shanghai

Source: https://english.news.cn/20260716/b0449aa2133542868e310fdc45ef2969/c.html · Xinhua · 2026-07-16 Verification: T2 verified · Xinhua primary + CGTN · policy-regulation

Update to the July 17 bulletin, which reported WAICO as “a Chinese diplomatic proposal” with “no participating countries or charter published.”

On July 16, representatives from 29 countries signed the founding agreement for the World Artificial Intelligence Cooperation Organization (WAICO) in Shanghai. Chinese Foreign Minister Wang Yi signed on behalf of China; UN Secretary-General António Guterres attended the ceremony. Named founding members include Russia, Kazakhstan, Pakistan, Indonesia, and Laos; other signatories span Africa, Latin America, and Central Asia, with Brazil and Venezuela among those cited. WAICO is established as an independent intergovernmental organization headquartered in Shanghai, mandated to “promote international cooperation and global governance on AI,” ensuring it is beneficial, safe, and fair. Xi Jinping’s keynote at the WAIC 2026 opening on July 17 called WAICO “an important milestone in the history of AI development” and pledged 5,000 AI training opportunities to developing nations.

Why it matters for automation/productivity: WAICO is now a functioning intergovernmental body — not a proposal — with 29 founding members largely from the Global South and China-aligned nations. For BD teams operating in ASEAN, MENA, or Sub-Saharan Africa, WAICO creates a second governance track for AI rules that will diverge from EU AI Act and U.S. executive order frameworks in material ways. Countries in China’s AI infrastructure orbit (those already using Kimi K3, DeepSeek, or state-subsidized compute) are more likely to adopt WAICO compliance requirements than EU/US frameworks. The practical compliance implication: for cross-border AI deployments in WAICO member countries, watch for WAICO governance requirements to emerge in procurement and data sovereignty requirements within 12–24 months.

Key claims:

• 29 countries signed WAICO founding agreement in Shanghai on July 16 → Xinhua (T2, 2026-07-16)
• Chinese FM Wang Yi signed for China; UN SG Guterres attended → Xinhua (T2)
• Named founding members include Russia, Kazakhstan, Pakistan, Indonesia, Laos → Xinhua (T2)
• WAICO mandate: independent intergovernmental organization, HQ in Shanghai → Xinhua (T2)
• Xi Jinping keynote July 17: called WAICO “an important milestone”; pledged 5,000 AI training opportunities to developing nations → CGTN (T3, 2026-07-17)

Caveats: Full list of 29 founding countries has not been published in English-language primary sources as of this run; named members are from Xinhua reporting. WAICO’s charter, decision-making procedures, and enforcement mechanisms have not been published. Xi’s speech text cited from CGTN (T3) reports, not a primary Xinhua transcript. The July 17 bulletin’s assessment (“no participating countries or charter published”) was accurate at its publication time; the signing ceremony occurred on July 16 after that run’s research pass.

Dropped

Items considered but not published, with reason.

Limitations

• Sources unreachable: TechTimes article on China companion rules returned HTTP 403; Axios AI regulatory convergence article returned HTTP 403. Content attributed to these sources is drawn from search result excerpts and corroborated by separately fetched secondary sources (IAPP, AI Governance Institute). x.ai/news returned intermittently restricted responses.
• Login-walled coverage: X/Twitter timelines, LinkedIn, Discord not directly accessible. Mandatory social search pass conducted per Stage 3 protocol; @AnthropicAI, @GoogleDeepMind, @Kimi_Moonshot, @huggingface handles searched via search-indexed posts. No X-primary items surfaced in this run.
• China agent governance primary text: The Implementation Opinions on Intelligent Agent Governance were not fetched from Chinese government sources (cyberspace.mii.gov.cn); English-language coverage from IAPP (T2) and AI Governance Institute (T3) was used. Translations and summaries may not capture the precise scope of sector definitions or filing requirements.
• WAICO full member list: All 29 founding countries not enumerated in English-language primary sources; only a subset (Russia, Kazakhstan, Pakistan, Indonesia, Laos, Brazil, Venezuela) are specifically named in Xinhua reporting.
• Hugging Face breach scope still open: The company noted partner/customer data exposure assessment is ongoing; the final impact scope and attack attribution have not been published.
• Categories with thin coverage: workflow-automation — no new end-user agentic product launch confirmed in window. agent-framework — no significant new framework version confirmed in window (LangChain 1.3.14 was a minor patch only). mcp-ecosystem — no new MCP servers or spec updates in window (final spec targets July 28). model-release — Gemini 3.5 Pro remains unconfirmed; no other frontier model release in window.
• Geographic bias: Stage 3.5 Indonesia/SEA search found no AI product launches from SEA-originated vendors in window. World AI Show Jakarta was July 7–8, outside window. WAICO membership includes Indonesia, which is covered in WAICO item.
• Vendor benchmark caveats: No new model benchmarks published in this window; benchmark caveats not applicable.
• Exploration note: 50% of search passes targeted non-registry sources (Hugging Face security blog, China AI Governance Institute, IAPP, Xinhua/CGTN, adversarial checks on Gemini Notebook, Hugging Face breach). Four of seven published items came entirely or primarily from non-registry sources.

Search log (compact)

Q: “AI announcement model release July 18 2026” → 9 results, 1 high-relevance [exploratory]
Q: “Anthropic Claude news July 18 2026” → 10 results, 4 in-window [registry]
Q: “OpenAI announcement July 18 2026” → 7 results, 0 July 18 confirmed [registry]
Q: “Google DeepMind Gemini news July 18 2026” → 8 results, Gemini 3.5 Pro still unconfirmed [registry]
Q: releasebot.io/updates/anthropic (web_fetch) → no July 18 updates; last Jul 17 = Claude Code 2.1.212 [registry]
Q: anthropic.com/news (web_fetch) → last post Jul 14; no Jul 15–18 posts [registry]
Q: “agent framework AI launch July 2026 release” → 9 results, 2 in-window candidates [exploratory]
Q: “MCP Model Context Protocol news July 2026” → 9 results, no in-window (RC published May 21; July 28 final) [registry]
Q: “AI startup funding launch July 18 2026” → 10 results, no confirmed July 18 [exploratory]
Q: “site:x.com AI announcement July 18 2026” → 10 results, WAIC + Kimi social posts [social - mandatory Stage 3]
Q: “WAICO founding agreement signed 29 countries July 17 18 2026” → 10 results, 8 high-relevance [exploratory]
Q: releasebot.io/updates/anthropic/claude-code (web_fetch) → 2.1.212 July 17 confirmed [registry]
Q: “Hugging Face new model release July 17 18 2026” → 10 results, Inkling on HF; HF security incident [exploratory]
Q: github.com/anthropics/claude-code/releases (web_fetch) → 2.1.210–2.1.212 confirmed; 2.1.212 July 17 [registry]
Q: blog.modelcontextprotocol.io (web_fetch) → no Jul 15–18 posts; last post Jun 29 [registry]
Q: “Meta AI Llama xAI Grok Mistral announcement July 18 2026” → 8 results, no in-window [registry]
Q: google.com/blog Gemini Notebook NotebookLM July 16 2026 [registry]
Q: blog.google/…/notebooklm-gemini-notebook/ (web_fetch) → T2 primary confirmed; July 16 [registry]
Q: 9to5google.com NotebookLM article (web_fetch) → Gemini 3.5 + Antigravity upgrade to AI Pro confirmed [T2-T3]
Q: techcrunch.com Gemini Notebook article (web_fetch) → confirmed rebrand details [T2-T3]
Q: ghacks.net Gemini Notebook + Cloud Computer AI Pro article (web_fetch) → Cloud Computer AI Pro expansion confirmed [T2-T3]
Q: “Hugging Face 'AI-agent' intrusion breach July 2026” → 9 results, 8 high-relevance [exploratory]
Q: huggingface.co/blog/security-incident-july-2026 (web_fetch) → T2 primary disclosure confirmed July 16 [exploratory - non-registry]
Q: “China AI agent regulations enforceable July 15 2026” → 10 results, 7 high-relevance [exploratory]
Q: iapp.org China companion rules article (web_fetch) → T2 confirmed; July 15 effective date [exploratory]
Q: “Altman Amodei Hassabis AI regulatory framework July 16 2026” → 10 results, 8 high-relevance [exploratory]
Q: “Anthropic hires misuse prevention analysts July 15 2026” → 9 results, 7 high-relevance [exploratory]
Q: “Gemini Notebook NotebookLM criticism concerns July 2026” → 9 results [adversarial]
Q: langchain-ai/langchain GitHub releases (web_fetch) → 1.3.14 July 16, minor patch only [registry]
Q: “LangChain LlamaIndex browser-use OpenHands July 2026” → 9 results, no in-window significant releases [registry]
Q: “AI Indonesia SEA startup AI lokal July 2026” → 9 results, World AI Show Jakarta Jul 7-8 only [geo-diversity Stage 3.5]
Q: “GitHub trending AI repositories week July 18 2026” → 9 results, no confirmed in-window item [Stage 3.5]
Q: “PulseMCP new MCP servers July 2026” → 9 results, last entries July 9 [registry]
Q: huggingface.co/papers (web_fetch) → SEED, SearchOS-V1 papers Jul 17; no near-term BD application [Stage 3.5]
Q: “site:x.com AI model release announcement July 17 18 2026” → 10 results, Gemini Notebook X post found [social - Stage 3]

Total searches: 35, of which 17 exploratory, adversarial, geo-diversity, or social passes (49%).

Suggested next runs

• Hugging Face breach final scope (Jul 19–21) — fetch huggingface.co/blog for a follow-up post with final impact assessment and attack attribution; monitor for any partner/customer data exposure disclosure.
• Gemini 3.5 Pro GA (when it ships) — watch ai.google.dev/gemini-api/docs/changelog; Pro rollout of Gemini 3.5 + Antigravity upgrade in Gemini Notebook also pending; these are connected.
• Kimi K3 open weights release (Jul 27) — Moonshot stated July 27 open-weight release; watch Hugging Face for availability; model card with sparsity ratio and full architecture should be published.
• MCP 2026-07-28 final spec — final spec ships July 28; stateless core, Tasks extension, and deprecated Roots/Sampling/Logging will be confirmed. Follow with SDK tier-1 adoption timeline.
• WAICO charter and governance rules — founding agreement is signed but operational charter not published; watch Xinhua and english.www.gov.cn for the governance structure and initial working groups.
• China AI Agent Governance filing deadlines — the regulation requires filings from operators in sensitive sectors; Chinese-language government sources should publish guidance on filing procedures and deadlines.
• Inkling benchmark scores — Arena.ai and Artificial Analysis scores pending since launch; agentic tool-calling results specifically not covered at launch.