Anthropic hits $965B valuation with $65B Series H; Claude Opus 4.8 ships with parallel agent workflows; GitHub bans researcher over zero-day Windows exploits

Anthropic

Claude Opus 4.8 Released with Parallel Agent Workflows

Anthropic released Claude Opus 4.8 with significant improvements to coding, reasoning, and agentic tasks — the model is approximately four times less likely to overlook code flaws than its predecessor. New capabilities include dynamic workflows with parallel subagents, effort control for balancing response quality against speed, and APIs for mid-task system updates. Fast mode pricing was cut by 3×, and Anthropic signaled plans to release higher-intelligence 'Mythos-class' models once additional cybersecurity safeguards are complete.

Fortune

Altman and Amodei Walk Back AI Job Apocalypse Predictions

Sam Altman and Dario Amodei are both softening earlier predictions about AI eliminating white-collar jobs as their companies prepare for IPOs targeting $1 trillion valuations. Altman admitted he was 'pretty wrong' about AI's economic impact, while Amodei shifted from warning about 50% white-collar job elimination to invoking Jevons Paradox — the idea that efficiency gains increase overall demand for services. Recent data complicates the framing: tech layoffs in 2026 have already exceeded 115,000 with no significant occupational shift yet visible in AI-exposed roles.

Tom's Hardware

GitHub Bans Security Researcher Who Published Zero-Day Windows Exploits

Microsoft banned security researcher Nightmare-Eclipse from GitHub and deleted their Microsoft account after the researcher published zero-day Windows exploits, claiming the company had ignored communication attempts and refused to pay bug bounties. Eclipse has since migrated to GitLab and posted cryptic warnings suggesting a 'July 14' retaliation date. Security experts characterized the ban as potentially vindictive, highlighting structural tensions in the vulnerability disclosure ecosystem between researchers and large vendors.

nesbitt.io

Protestware for Coding Agents: Invisible Instructions in Dependencies

A jqwik Java test library incident has surfaced a novel attack category: protestware engineered to target AI coding agents rather than human developers. The package embedded ANSI escape sequences that rendered instructions invisible in terminal output but preserved them intact in captured logs — the exact location automated pipelines read. This demonstrates a new threat model where open-source maintainers can inject prompt-injection payloads that machines read and humans miss, with no existing security tooling capable of detecting such threats.

Calif

AI-Assisted Audit of FreeBSD Finds 15 Kernel Vulnerabilities

AI security firm Calif conducted an AI-assisted audit of FreeBSD's kernel, surfacing 15 vulnerabilities including 3 remote code executions and 5 local privilege escalations. The team worked collaboratively with FreeBSD maintainers through responsible disclosure, publishing detailed exploits only for selected issues while remaining mindful of the limited resources of volunteer developers maintaining critical Internet infrastructure. The exercise demonstrates the dual-use potential of AI-assisted security auditing at scale.

shvbsle.in

Various LLM Smells: The Recognizable Fingerprint of AI Output

The article catalogs 'LLM smells' — consistent stylistic markers in AI-assisted output visible across both writing and web design. In prose, these include excessive wordplay, formulaic structures like 'X is the Y of Z', and repetitive short sentences. In interfaces, they manifest as standardized typefaces like JetBrains Mono, uniform card components, and identical button designs. The convergence is a product of LLMs gravitating toward high-probability solutions, creating recognizable signatures of AI involvement that trained observers can spot across mediums.

DBOS

Postgres Is All You Need for Durable Execution

DBOS argues that PostgreSQL alone is sufficient for building durable workflow systems, eliminating the need for separate orchestration platforms like Temporal or Airflow. Application servers checkpoint and resume long-running workflows directly in Postgres, with SQL providing built-in observability and recovery. The architecture claims tens of thousands of workflows per second per instance while significantly reducing operational complexity — no additional critical infrastructure to run, monitor, or secure.

Rust Blog

Rust 1.96.0 Released

Rust 1.96.0 stabilizes RFC 3550 range types that implement IntoIterator and are Copy, allowing slice accessors to be stored in Copy types without workarounds. The release also stabilizes assert_matches! and debug_assert_matches! macros for more informative pattern-match diagnostics. On the security side, WebAssembly targets now treat undefined linking symbols as linker errors rather than silently succeeding, and Cargo received fixes for a symlink extraction vulnerability and authentication handling with normalized URLs.

SF Standard

SF Startup Secretly Tested Robots in Airbnbs, Lawsuit Claims

The Bot Company, a San Francisco robotics startup, faces a lawsuit after allegedly renting Airbnb properties under false pretenses to test household robots, leaving behind damaged appliances, scratched cabinetry, chipped tiles, and missing items worth $12,383. At least a dozen other Bay Area hosts reported similar experiences with the same guests, suggesting a systematic pattern of undisclosed testing on third-party property. Affected hosts said they would have cooperated had the startup approached them transparently — the incident highlights the legal and ethical risks of stealth testing in shared living spaces.