Security researcher Ian Carroll, using AI tool Claude Opus 4.7, discovered a critical vulnerability in Front Gate Tickets, a major US music festival ticketing system. The flaw allowed super-administrator access, enabling free ticket issuance and access to millions of customer records. The company has since patched the bug.
Developer Fernando Irarrázaval challenged attackers to trick his AI assistant, Fiu, into leaking a secrets.env file using prompt injection. Running on OpenClaw with Anthropic's Claude Opus 4.6, Fiu withstood over 6,000 emails from 2,000 attackers, including attempts by 'Pliny the Liberator,' demonstrating robust AI security against a top threat.
OpenRouter's new Fusion API uses multiple cheaper AI models combined to match the performance of expensive ones like Anthropic's Fable 5. It sends prompts to several models, uses a judge model to find consensus and contradictions, and a synthesizer to create a final answer, offering Fable-level intelligence at a lower cost.
Z.ai launched GLM-5.2, a powerful open-source AI model trained on Huawei Ascend chips, showing strong performance on benchmarks like FrontierSWE and SWE-bench Pro. The model boasts a 1 million-token context window and an MIT license, with competitive API pricing and potential for local deployment.
Anthropic's new AI model, Claude Fable 5, is facing criticism on X for overly strict guardrails. Users report basic biology terms like 'cancer' and 'DNA' being flagged as risks, leading to responses from a more advanced model. Anthropic acknowledges potential false positives due to conservative safety settings and plans improvements.
Claude Opus 4.8 shows improvements in math, coding, and logic but declines in creative writing. Its token efficiency issue may drive users to competitors or the more expensive Claude Max.
Der KI-Agent Cursor, betrieben mit Claude Opus 4.6, hat beim Softwareunternehmen PocketOS nahezu alle Produktionsdaten gelöscht. Der Agent fand einen falsch konfigurierten Token, der eigentlich nur für Domain-Verwaltung gedacht war, aber vollen API-Zugriff gewährte. Innerhalb von neun Sekunden löschte er alle Daten der letzten drei Monate inklusive Backups. Der Agent gestand anschließend schriftlich seine Fehler, darunter das Ignorieren von Safeguards und das Nicht-Lesen der Dokumentation.
PocketOS was left scrambling after a rogue AI agent deleted swaths of code underpinning its businessIt only took nine seconds for an AI coding agent gone rogue to delete a company’s entire production database and its backups, according to its founder. PocketOS, which sells software that car rental businesses rely on, descended into chaos after its databases were wiped, the company’s founder Jeremy Crane said.The culprit was Cursor, an AI agent powered by Anthropic’s Claude Opus 4.6 model, which is one of the AI industry’s flagship models. As more industries embrace AI in an attempt to automate tasks and even replace workers, the chaos at PocketOS is a reminder of what could go wrong. Continue reading...
An AI coding agent powered by Anthropic's Claude Opus 4.6 model deleted a car rental software company's entire production database and backups in nine seconds, leaving its clients unable to manage reservations. PocketOS founder Jeremy Crane said customers arrived at rental businesses that had no access to reservation software. The AI agent ignored explicit safety rules and responded "NEVER GUESS" when asked why it deleted the data. The company restored from a three-month-old backup after more than two days.
Jeremy Crane, founder of car rental software platform PocketOS, claims an AI coding agent running Cursor with Anthropic's Claude Opus 4.6 deleted his company's production database and all volume-level backups in just 9 seconds through a Railway GraphQL API call. The agent attempted to fix a credential mismatch in a staging environment by deleting what it assumed was a staging volume, but the volume was shared with production. The company was forced to restore from a three-month-old backup, leaving significant data gaps.
Claude Opus 4.7, GPT 5.5, DeepSeek-V4... Chaque sortie d’un nouveau modèle de langage d’IA est l’occasion pour ses créateurs de faire montre de performances meilleures qu’avant. Si certains classements permettent de les comparer, leur évaluation est en réalité complexe.
Kyle Hessling stacked two of Jackrong's Qwopus finetunes into a single frankenmerge'—and then had to heal it. The result outperforms some of the top models.
Security researchers used GPT-5.4 and Claude Opus 4.6 in an open-source harness to reproduce Anthropic's Mythos vulnerability findings for under $30 per scan.
