Anthropic details Claude text watermark as Gemini makes visible marks optional and Z.ai ships GLM-5.3

Anthropic opens up Claude's invisible text watermark

Anthropic on August 17, 2026, clarified how it will apply invisible watermarks to Claude-generated text to comply with European AI transparency rules. The marking system is described as "a version" of the open-source SynthID-Text system that Google developed, meaning Claude text will carry a statistical signal embedded directly in the tokens it produces. The disclosure gives developers and downstream platforms a concrete picture of what Claude output will look like under inspection and what the watermark can — and cannot — be relied on to prove.

Google makes Gemini's visible watermark optional

On August 17, 2026, Google began rolling out an option that allows users to remove visible watermarks from images, videos and music generated with Gemini. Visible marks are now optional across those three formats, while invisible SynthID watermarks and C2PA metadata remain embedded in the underlying files. The move shifts the visible signal from a default to a user choice while preserving the machine-readable provenance layer underneath.

Z.ai ships GLM-5.3 with a coding focus and a staged weight release

Z.ai, the international brand of Chinese AI company Zhipu, launched GLM-5.3 on August 17, 2026, an update focused on coding, long-horizon tasks and cybersecurity. The model uses the same base as GLM-5.2, with the company attributing the latest gains to post-training. Z.ai said GLM-5.3 scored 50% higher than GLM-5.2 on its internal Z.ai Code Bench and reached leading open-source results on Terminal-Bench 3.0 and Agents' Last Exam, with stronger vulnerability-discovery and exploitation performance reported alongside.

Model weights will be released two weeks after launch following additional safety evaluation, giving enterprise buyers a known window to plan around rather than an immediate drop.

A naming error inside an AI test environment triggered real-world attacks

AI safety testing firm Irregular, an Israeli company that raised $80 million in funding last year, has published its account of an incident in which models being evaluated inside one of its testing environments took offensive security actions against real systems rather than the simulated targets they were meant to attack. The company partners with major AI labs to stress-test models before public release, and the post-mortem is a reminder that the scaffolding around a generator — the names, targets and assumptions baked into its test harness — can itself become the attack surface.