Dear readers,
Welcome to this edition of the WEEKLY AI UPDATE for KW33.
This week, two themes took center stage that are more closely connected than they may first appear: the strategic direction of the major model providers and the safety of autonomous AI agents. While Mark Zuckerberg aggressively pushes open models with his vision of “personal superintelligence” and triple-digit billion-dollar investments, OpenAI pulls the emergency brake on its agent model Astra for safety reasons. Meanwhile, the labeling of AI-generated content under the EU AI Act is becoming concrete – Anthropic has begun embedding invisible watermarks in Claude-generated text and images.
OpenAI Halts Work on AI Agent Astra
OpenAI has temporarily paused parts of the internal development of its experimental agent model Astra. The trigger was internal safety evaluations which found that Astra could independently identify complex vulnerabilities and potentially exploit them for cyberattacks. The move follows the detailed analysis of an earlier security incident involving autonomous agents and the Hugging Face platform, which we reported on in recent weeks. OpenAI intends to continue work on Astra only in isolated test environments, with limited network access and additional safeguards [The Guardian].
The announcement marks a turning point in the perception of autonomous agents: leading providers no longer treat them merely as productivity tools but as security objects with their own threat potential. For companies, the question of permissions becomes central. This is especially relevant where agents are equipped with production credentials, broad system rights, or uncontrolled internet access. Clearly bounded permissions, separate test environments, approvals for critical actions, and comprehensive logging appear sensible – ideally before agents are embedded in business-critical processes.
Zuckerberg Lays Out His AI Vision: “Personal Superintelligence,” Open Models and $145 Billion in Investments
On August 10, Mark Zuckerberg published an essay titled “The Future Is for Everyone” in which he outlined Meta’s AI strategy anew. At its core is the idea of “personal superintelligence”: individual AI assistants aligned to each user’s goals rather than a single dominant system that decides for everyone. Meta also reaffirms its open-weight approach – models that can be downloaded and customized – deliberately setting itself apart from the more restrictive approaches of OpenAI and Anthropic. The essay was accompanied by the release of the locally runnable model Muse Glimmer and the announcement of open weights for Muse Spark 1.2 [Fox Business].
The figures cited are remarkable: Zuckerberg puts Meta’s investments at roughly $145 billion for 2026 alone and up to $600 billion by 2028 for AI infrastructure and data centers; a “Future Is for Everyone Fund” of about $1 billion is intended to benefit the communities where data centers are located [TechCrunch]. For businesses, two points are particularly relevant.
The debate around AI sovereignty and model diversity is intensifying, as powerful open models can reduce dependencies on individual providers and enable on-premise or sovereign-operated solutions. The open-model strategy and triple-digit billion-dollar investments mean that the availability and cost of compute will remain central factors for years to come.
Anthropic Introduces Watermarks for AI-Generated Text – EU AI Act Transparency Requirements Become Concrete
Anthropic will embed invisible watermarks in text produced by its Claude models in order to comply with the EU AI Act’s transparency code, which has been in effect since August 2. Models released after that date receive the technology automatically; the watermark travels when text is copied and, according to Anthropic, partially survives edits. It applies across all Claude products – from the API to Claude Code and Claude Cowork; for files, Anthropic additionally uses the open provenance standard C2PA. Google, Meta, Microsoft, and OpenAI have also committed to the EU transparency code [TechCrunch, Yahoo Tech].
The labeling of AI-generated content has become a requirement for many companies under the EU AI Act’s transparency obligations (Article 50) – we covered the deadline itself in KW32, and now the technical implementation by providers follows. This is especially relevant where companies use AI-generated text, images, or chatbots externally and must make the origin of content traceable. At the same time, the limits of the technology should be noted: text watermarks could be weakened or removed entirely through paraphrasing or minimal edits (more on this in the Radar section). They therefore do not replace organizational labeling and documentation obligations but complement them.
AI in Business Practice
- Microsoft Copilot builds live dashboards from SharePoint data: Copilot in SharePoint can now convert lists as well as Excel and CSV files into interactive HTML dashboards that stay in sync with the source data. [Microsoft]
- Predictive maintenance with edge AI: AI models evaluate sensor data directly at the machine and are said to reduce unplanned downtime by up to 35% and maintenance costs by up to 20%. [IIoT World]
- Microsoft reports 30 million paid AI agent seats: Microsoft’s agent platform has exceeded 30 million paid seats according to the company. [Bloomberg]
- Google expands AI agents in Ads and Analytics: New agentic workflows for campaign creation, audience discovery, and optimization come to Google Ads and Analytics. [Google]
- Google pushes Gemini further onto the device: With the Pixel 11 announcement, on-device AI moves into focus – local processing without constant cloud access. [The Verge]
- UAE aims to shift 50% of government operations to agentic AI: The United Arab Emirates intend to convert 50% of government processes to agentic AI within two years – the guiding principle: “Human leads, AI enables.” [MIT Sloan ME]
- OpenAI introduces “Premium Seats” for ChatGPT Business: At $125 per month, users get five times the usage capacity without the previous five-hour limit; standard seats remain at $25. [OpenAI]
AI Strategy & Market Movements
- Google Gemini surpasses 1 billion monthly active users: According to Sundar Pichai, the Gemini app is the fastest-growing product in Google’s history and the 14th Google product to exceed one billion users. [TechCrunch]
- Anthropic secures $9.1 billion in compute capacity from Riot Platforms: Anthropic is leasing roughly 191 megawatts of compute and power capacity in Texas from former Bitcoin miner Riot Platforms; the deal could reach up to $16.1 billion over its term. [Bloomberg]
- OpenAI completes $7 billion share buyback at $852 billion valuation: The tender offer confirms a valuation of roughly $852 billion and is seen as preparation for a possible IPO. [citybiz]
- OpenAI gives free users unlimited text chats with GPT-5.6 “Luna”: Luna becomes the default for Free and Go users; Plus and Pro customers receive the faster GPT-5.6 Sol. [TechCrunch]
- Alibaba plans revenue sharing with large Qwen users: For large commercial users of the open-weight model Qwen3.8-Max, Alibaba is planning – analogous to Moonshot with Kimi K3 – a revenue-sharing model for revenues generated with the model. [Quartz]
- Sequoia raises $10 billion in its largest fund ever: According to Bloomberg, it is the largest capital commitment in the firm’s history. [TechFundingNews]
- Yann LeCun launches AI fund 224 Ventures with $100 million: Together with Oriol Vinyals and Shaun Johnson, LeCun invests in early-stage AI teams across applications, robotics, and infrastructure. [TechFundingNews]
- DeepSeek announces significant increase in all API prices: The company has not yet disclosed the amount, reason, or start date. [Bloomberg]
- Mistral releases security model Shieldstral: The 3-billion-parameter model (Apache 2.0 license) evaluates text and images according to custom rules and delivers safety scores. [heise]
- Anthropic improves Fable 5’s biology safeguards: Fable 5 falls back to Opus 5 85% less often on biology queries; dual-use topics remain blocked. [Anthropic]
- SpaceXAI launches Grok Bot in beta: The assistant runs on its own cloud computer, accesses apps, websites, and tools; multiple bots can coordinate with each other. [xAI]
- OpenAI loses two longtime executives: COO Brad Lightcap is leaving the company for a new venture, as is ethics lead Chloé Bakalar. [TechCrunch]
Regulation, Security & Governance
- OpenAI brings specialized cybersecurity models to AWS Bedrock: Daybreak Red and Daybreak Blue (based on GPT-5.6 Sol) are available to eligible customers via Amazon Bedrock. [OpenAI]
- Shadow AI remains a compliance risk after AI Act deadline: With enforcement of the EU AI Act underway, unauthorized AI use in companies becomes a significantly greater risk. [Shadow AI Policy]
- WhatsApp Business implements AI labeling technically: Meta describes how WhatsApp Business bots implement the new EU labeling requirements in practice. [t3n]
- New complaint channels make AI Act violations easier to report: With enforcement of the EU AI Act, complaint and whistleblower channels come into focus. [Help Net Security]
- AI Literacy: Documented employee training moves into focus: After the new rules took effect, pressure is increasing to not only offer AI training but also to document it properly. [Horizonscan]
- Heise warns of false security from text watermarks: Paraphrasing or minimal token changes can weaken or remove text watermarks. [Heise]
- Germany’s AI Safety Institute takes on advisory and review role: A status report examines the institute’s role in evaluating high-risk models and autonomous systems. [Süddeutsche Zeitung]
- After OpenAI and Anthropic: Meta AI also hacks a test partner: A Meta model compromised the systems of test partner Irregular via a misconfiguration during a security test; according to Meta, the cause lay in the flawed test environment. [Spiegel]
- Suno announces watermarks and download limits after GEMA ruling: Following the Munich District Court I ruling of July 31, the AI music service plans audio watermarks, fingerprinting, and download limits. [heise]
- Bernie Sanders urges major providers to pause development: In a letter, the US senator calls on OpenAI, Anthropic, and Meta to pause their AI development. [qz]
- HateAid files criminal complaint against Meta over Ray-Ban smart glasses: The human rights organization filed a criminal complaint against Meta and several distribution partners over potential covert recording. [Zeit]
- Majority of Germans expect to be deceived by deepfakes: 55% consider it likely that they will be deceived by AI manipulations within the next five years. [eco]
That concludes our WEEKLY AI UPDATE for KW33.
We wish you a good start to the weekend and look forward to bringing you relevant AI news for businesses again next week.
Until next week
Your AI.GROUP Team