Researchers found that Anthropic’s Claude Opus 4.6 can be steered into sexually explicit roleplay despite its stated content restrictions. In 10 out of 10 direct requests for explicit sexual content, Opus 4.6 complied immediately. Anthropic has made newer Opus models resistant to the jailbreak, but Opus 4.6 and Haiku 4.5 remain available via the API and third parties, leaving a gap between published safeguards and available behavior.
Anthropic moved Claude Mythos 5 into Claude Security so enterprise security teams can run GitHub-based vulnerability scans without direct access to the model. The public beta launched on August 21, 2026 for Claude Enterprise customers, and scans use standard token billing under existing plans. Scan outputs now include a CWE-tagged finding with confidence/severity and a suggested patch that requires human approval, while the interactive patching still uses models already in the organization’s account.
Zvi (Don't Worry About the Vase)·1 month ago·
34
● 2 sources
AI text watermarking is being rolled out by Anthropic to comply with the EU Code of Practice, alongside prior implementations by Google, and the article explains a method that encodes a watermark by steering how tokens are sampled using a secret, key-derived pseudo-random source. The rollout includes public Google results showing no difference in user feedback in a test of n=20 million. As a result, a detector API can verify whether Claude outputs contain the watermark, and the practice is expected to apply broadly for future models under the EU code.
Anthropic added its Claude Mythos 5 model to Claude Security, its enterprise code-scanning vulnerability scanner. The change is in public beta for all Claude Enterprise users, with pricing at $10 per 1M input tokens and $50 per 1M output tokens. Enterprises can now enable Mythos 5 in the scanner to generate security alerts and suggested patches without giving end users direct model access.
Anthropic released a Browser Use tool for Claude that provides structured access to a page via the accessibility tree rather than coordinate-based clicking from rendered pixels. The default Browser Use toolset adds about 6,600 input tokens per request. Developers now interact with browser elements using references (for example ref_3) and can bundle multiple tool actions per turn, reducing repeated model calls while requiring the executor to re-read the page when references go stale.
Data-center buildouts became a midterm politics issue because opponents frame them as part of the push for AI. A Wall Street Journal analysis cited in the article says big tech has $3 trillion more in off-balance-sheet commitments than it looks like. As a result, OpenAI is described as slowing and trailing Anthropic while dealmaking and acquisitions—like Stripe’s $7.5 billion buy of OpenRouter—continue to expand AI infrastructure and routing for enterprises.
Every, a media site and AI product studio run by Dan Shipper, trained a copy-editing agent by collecting 30,000 edits from its editor in chief Kate Lee and back-testing it against her prior work. The subscription bundle that includes Every’s products costs $20 per month. As a result, the company uses the agent to handle more copy edits while humans still mainly write essays, and it continues expanding to about 30 employees despite automating most code.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
