Anthropic: Claude Opus 5.5 zeigt 85 % weniger erfolgreiche Grenzumgehungen

·3 views

Anthropic hat Claude Opus 5.5 am 22. September 2026 veröffentlicht und verweist dabei auf 85 % weniger erfolgreiche Versuche der Grenzumgehung im Vergleich zum Vorgänger. Die Angabe ist relativ und stammt aus einem Labor-Testaufbau, sodass etwa 15 von 100 früher erfolgreichen Versuchen weiterhin gelingen. Der Artikel warnt davor, diese Modelleigenschaft mit der Sicherheit einer konkreten API gleichzusetzen, da Prompts, Tool-Schemas und Retrieval-Quellen nicht erfasst werden. Empfohlen wird, die tatsächlichen Funktionsaufrufe des Agenten mit einer adversarialen Testsuite zu prüfen, die laut Autor nur einen einstelligen Dollarbetrag pro Lauf kostet. Das Modell bietet zudem ein Kontextfenster von 1 Mio. Tokens und kann über 18 Stunden an einer Aufgabe arbeiten.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Quick Code Review Approvals May Be Hiding Real Problems

A common pattern in software teams is a large pull request receiving approval within minutes, with no comments or questions. Such rapid sign-offs can leave real defects unnoticed until they reach production. The author suggests that reviewers may be busy, trust the author, or hesitate to challenge more senior colleagues. Authors can help by keeping changes smaller, adding a clear description, and flagging specific areas where they want scrutiny. Thanking reviewers who raise concerns and examining one's own review habits are also recommended to shift the team away from treating review as a formality.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Lixity Adds Contextual Tense Explanations and Speech-Plan Export to Manuscript Review

The article describes Lixity's manuscript and research review workflow, which adds optional grammar models for tense analysis and a Qwen3-TTS worker for listening profiles. It explains how passage-anchored speech directions can be exported as JSON or SSML. The piece argues that a marked paragraph is only useful when the software shows the reason for the mark, since tense-pattern counts are heuristic and need an author's or editor's judgment. It also describes how research panels should follow the reader's current passage, retiring stale context on chapter changes and keeping the return control simple.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Claude Opus 5.5: el 85 % de reducción en elusiones no equivale a seguridad de API

Anthropic lanzó Claude Opus 5.5 el 22 de septiembre de 2026 y reportó un 85 % menos de intentos exitosos de elusión de límites frente a su predecesor. Esa cifra relativa proviene de pruebas de laboratorio sobre el modelo, por lo que no refleja el comportamiento de prompts, esquemas de herramientas o fuentes de recuperación concretas de cada aplicación. Un artículo de DEV Community advierte que los equipos no deben interpretarla como una resolución de la inyección de prompts. En su lugar, recomienda verificar las llamadas que hace el agente y ejecutar suites adversariales propias para medir el riesgo residual en cada API.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DevOps Engineer Builds Self-Hosted ChatGPT-Style Platform on a Laptop

A DevOps engineer built OzAI, a self-hosted AI platform running open-weight models Qwen3 8B and Qwen2.5-Coder 3B on a MacBook Pro M1 Pro with 16 GB of RAM. The setup uses Ollama for serving models, LiteLLM as an OpenAI-compatible gateway with per-user API keys and rate limits, and Open WebUI as a chat interface. Ollama runs natively on macOS rather than in Docker, since containers cannot access the Apple GPU. Benchmarks showed the 3B model generating about 53 tokens per second and the 8B model about 22, which the author considers sufficient for single-user use. The project is the first of a three-part series, with later parts covering observability and a move to GPUs on Amazon EKS.

0 comments Read more at DEV Community