Three fabricated documents were injected into a ChromaDB knowledge base, resulting in a RAG system inaccurately reporting a company's Q4 2025 revenue as $8.3M, a 47% decrease year-over-year, along with a planned workforce reduction. This process was completed in under three minutes on a MacBook Pro without GPU support or cloud services.

AI techniques can analyze burner accounts on social media to accurately identify pseudonymous users. Experiments show a higher success rate in correlating individuals with accounts across multiple platforms compared to traditional deanonymization methods.

Se compartió con Dropbox.

Anthropic's Frontier Safety Roadmap emphasizes the need for improved security measures to prevent theft and manipulation of AI models. The roadmap also focuses on implementing safeguards to prevent dangerous uses of AI and ensuring model alignment to avoid autonomous harm.

Gemini 3.1 Pro is the latest model in the Gemini 3 series, featuring advanced multimodal reasoning capabilities. Model cards provide essential information about the models, including limitations, mitigation strategies, and safety performance, and may be updated to reflect improvements.

AI agents are currently deployed in diverse contexts, ranging from email triage to cyber espionage. An analysis of millions of human-agent interactions across Claude Code and a public API aims to measure the autonomy of AI agents in real-world usage.

A new benchmark evaluates outcome-driven constraint violations in autonomous AI agents to enhance safety and alignment with human values. This benchmark addresses limitations of existing safety assessments that mainly focus on harmful actions.

Prompt injection in Google Translate can reveal the underlying instruction-following language model. Responses indicate that the model lacks strong boundaries between processing content and following instructions.

OpenScholar synthesizes scientific research and accurately cites sources, achieving a level of accuracy comparable to human experts. In a study, researchers found that the AI model significantly reduces the issue of hallucination seen in other models like GPT-4o, which fabricated 78-90% of its outputs.

Claude Opus 4.6 features significant advancements in AI models' cybersecurity capabilities. Experts believe the current moment is critical for accelerating the defensive use of AI in response to the increasing risk of LLM-discovered zero-day vulnerabilities.