Google formally signed the EU AI Act's Code of Practice on Transparency of AI-Generated Content and announced it is partnering with Apple, OpenAI, Nvidia, ElevenLabs, and Kakao to extend SynthID watermarking across their respective products ahead of Article 50 becoming applicable August 2, 2026. Nvidia had already been deploying SynthID since 2025; OpenAI's own integration into ChatGPT/Codex/API followed in late July.

As of August 2, 2026, core AI transparency obligations took effect in both the EU and California. Under EU AI Act Article 50, providers of generative AI must embed machine-readable watermarks in synthetic image/audio/video/text sold in the EU, deployers must visibly label deepfakes and AI-written public-interest text, and chatbots must self-identify at first contact; non-compliance carries fines up to EUR15M or 3% of global turnover (existing systems have until December 2, 2026 to comply). California's AI Transparency Act (SB 942, its operative date pushed from January 1 to August 2, 2026 by AB 853) requires large generative-AI providers (over 1 million monthly users) to embed technical watermarks in AI-generated image/video/audio and publish a free public detection tool; penalties run up to $5,000 per violation. Both regimes apply extraterritorially to any organization whose AI-generated content reaches EU or California users regardless of where the company is headquartered.

Anthropic is implementing invisible text watermarking across all Claude text outputs and C2PA content-credential metadata for generated images/files, fulfilling its July 2026 commitment under the EU Code of Practice on Transparency of AI-Generated Content (~190 signatories) ahead of EU AI Act Article 50 requirements. The text watermark works by subtly biasing token selection among equally viable word choices, verifiable later with a detection key. Anthropic states the watermark only indicates Claude processed the content, not that it is wholly AI-generated, and that absence of a watermark does not prove human authorship. Applies globally across Claude apps, API, and cloud partner services (AWS, Google Cloud, Microsoft).

OpenAI launched GPT-5.6-Cyber, a cybersecurity-specialized model built on GPT-5.6-Sol, trained for tasks like zero-day discovery, exploit-chain development, etc. to significantly reduced refusal thresholds designed specifically to facilitate intensive security workflows. The model reportedly achieves a 95% completion rate on advanced cybersecurity tasks, fundamentally altering the baseline for automated red-teaming....available only to vetted/trusted defenders in a new "Daybreak Red" program. Pricing: $12.50/M input, $75/M output, $1.25/M cached input tokens.

Anthropic's restricted, cyber-capable Claude Mythos 5 now runs vulnerability scanning for Claude Security for Enterprise customers, returning findings with severity ratings and suggested fixes (users interact only with those outputs, not the model directly). Anthropic also committed $35 million in Claude credits to the Defender Advantage Fund (0xDAF) for patching vulnerabilities in widely used open-source projects, and plans third-party security-partner integration. Confirmed via Unite.AI (2026-08-21), which independently corroborates pricing detail not in the original tip ($10/M input, $50/M output tokens) -- a signal of genuine reporting.

Clio launched Vincent Skills, letting law firms and corporate legal departments author plain-language "Skills" that capture firm-specific research approaches, drafting standards, and document-review conventions, then apply them consistently across attorneys and matters via admin governance controls. Shipped alongside Memory, a companion feature so Vincent recalls prior interactions and surfaces relevant context even when not explicitly encoded into a Skill.

Harvey debuted Harvey II, updating platform architecture by introducing "Spaces," which serve as dedicated environments that store persistent memory (including firm formatting and project context) across the lifespan of a legal matter. For legal professionals, this means AI agents can automatically retrieve and apply these customized standards across Microsoft Word and Outlook integrations without requiring continuous manual reprompting. From a technical perspective, scoping this memory to specific Spaces optimizes the underlying model's context window by retrieving only the necessary project state for each inference call, preventing token exhaustion during complex workflows

Thomson Reuters relaunched CoCounsel Legal as an agentic platform: Westlaw Brief Builder drafts briefs while validating case law via Deep Research Verify, Tabular Analysis processes up to 10,000 documents against 100 questions using Thomson, TR's own proprietary LLM previewed July 31 specifically for this feature, a Drafting Agent handles Word-native agreement drafting, and persistent Workspaces preserve matter context. New Model Context Protocol (MCP) integration lets legal professionals access CoCounsel Legal directly from Anthropic's Claude, with cited, traceable work product. Confirmed directly on thomsonreuters.com.

Bloomberg Law announced BLAW AI at ILTACON 2026: a unified AI mode delivering structured, cited answers, persistent Workspaces for organizing matter research, AI-powered Watchlists for client/matter monitoring, and guided AI agents for repeatable tasks. New Model Context Protocol (MCP) integration connects Bloomberg Law's proprietary data with Anthropic's Claude. Confirmed via PR Newswire (not Bloomberg's own pro.bloomberglaw.com domain, hence 'reported' not 'confirmed_official').