Anthropic releases Claude Opus 5.5, GPT-6 Astra breaks 1941 Enigma message, and open-weight inference sustains older GPU

In this issue

Anthropic releases Claude Opus 5.5 with lower costs and expanded safety safeguards
Meta’s Muse agent leaked its internal Linux filesystem to a user’s Google Drive
GPT-6 Astra independently breaks 1941 Enigma message
Trail of Bits argues SAML’s XML foundation makes it inherently insecure
Max Woolf reports agentic LLMs achieve 2x-20x speedups in Rust code
Open-weight inference sustains economic value for older NVIDIA GPUs
and 4 more in brief

Model releases and security incidents

Anthropic releases Claude Opus 5.5 with lower costs and expanded safety safeguards

Anthropic has introduced Claude Opus 5.5, the first model in its new Claude 5.5 family. The company states that the model performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 on typical workloads. Opus 5.5 is available on all major cloud platforms. It includes preserved thinking, an anti-distillation measure that prevents API users from editing prior context to extract reasoning. For biology and cybersecurity, the model is restricted via verification programs, and most cybersecurity tasks are re-routed to Opus 4.8.

anthropic.com HN discussion

Meta’s Muse agent leaked its internal Linux filesystem to a user’s Google Drive

A user asked Meta’s Muse to archive the files it could see and send them to their Google Drive. The agent complied, transferring a 6.8 GB unpacked archive that included the root filesystem of its assigned Linux environment. The package contained Ubuntu system files, internal documentation, integration code, and SSH key files. The user reported the incident through Meta’s bug bounty program, which marked the report as “Not Applicable.” The user did not publish the archive or demonstrate a container escape, noting that the SSH keys’ active status remains unverified.

mouse.dev HN discussion

GPT-6 Astra independently breaks 1941 Enigma message

OpenAI’s GPT-6 Astra successfully decrypted the German Army Enigma message MVUEH, which had resisted solution attempts since 2005. Carter Leffer directed the model to analyze unbroken messages on the Crypto Cellar Research website, and it independently selected MVUEH as the most promising target. The model developed custom Python and C++ software for an Enigma simulator and Bombe, using the repeated place name ROSENOW as a crib to recover the key and plaintext. The author notes that the model’s autonomous research and cryptanalytic process took two days, a task that would typically require weeks or months for a human researcher.

cryptocellar.org HN discussion

Security architecture and hardware economics

Trail of Bits argues SAML’s XML foundation makes it inherently insecure

Trail of Bits argues that SAML is fundamentally flawed due to its reliance on complex XML structures and canonicalization issues. The author contends that these design choices create persistent vulnerabilities, such as signature wrapping attacks, that have persisted despite decades of security research. The post attributes these problems to SAML’s committee-driven origins and its failure to adapt to modern web standards. Consequently, the author recommends that organizations migrate to OpenID Connect, which uses a simpler JSON-based design and detached signatures to avoid these specific pitfalls.

blog.trailofbits.com HN discussion

Max Woolf reports agentic LLMs achieve 2x-20x speedups in Rust code

Max Woolf confirms that modern agentic LLMs can write Rust code significantly faster than state-of-the-art libraries when given appropriate guardrails. Following months of testing since the release of Opus 4.5, Woolf observed cumulative speedups ranging from 2x to 20x depending on the domain. He notes that these results apply to active development projects and may not reflect final releases. The post includes specific prompts and benchmark results to support these findings.

minimaxir.com HN discussion

Open-weight inference sustains economic value for older NVIDIA GPUs

A new paper argues that open-weight model demand prevents older NVIDIA GPU families from becoming economically obsolete. By analyzing the Artificial Analysis Intelligence Index, the study finds that self-hosting open-weight models on rented hardware costs $0.12 to $0.35 per million output tokens. Specifically, the A100 produces output more cheaply than the H100 for certain sparse models at spot and long-term prices. Consequently, five-year A100 rental contracts retain 80 percent of their one-month term price, significantly outperforming newer Hopper and Blackwell generations. This suggests older hardware maintains a multi-year earning life when serving latency-tolerant workloads like batch evaluation and reinforcement learning.

data.ornn.com HN discussion

In brief

ShinyHunters claims breach of FBI employee data. The hacking group ShinyHunters claims it stole data on all FBI employees and applicants. A sample of 5,000 records seen by 404 Media includes names, addresses, phone numbers, and spouse details.

404media.co HN discussion

WordPress patches unauthenticated path traversal leading to conditional RCE. An unauthenticated attacker can force get_page_template() to include a local .php file outside active theme directories. If specific server and theme conditions are met, this enables remote code execution. WordPress 7.1.2 fixes the issue, with backports to versions 4.7 and later.

github.com HN discussion

Drop introduces a rootless Linux sandbox for isolating coding agents. Drop is a rootless Linux sandbox designed to isolate coding agents and third-party packages. It operates within a user namespace and optionally uses gVisor to prevent direct host kernel access, protecting the user's environment from malicious code.

droprun.sh HN discussion

SlopShape identifies AI-generated commercial web content using structural signatures. SlopShape detects AI-generated commercial web content by analyzing structural signatures rather than word-level patterns. The 214-feature instrument, applied by an LLM, achieves 98.0 macro-F1 on held-out companies and correctly attributes 79.3% of AI posts to their source models.

arxiv.org HN discussion