Articles by tag: ai-safety

Fears of AI: why leaders are changing their rhetoric

After attacks on OpenAI's CEO, the company calls for restraint. However, it was their apocalyptic predictions that provoked aggression. Analysis of the situation.

AI agents ignore instructions: rise in incidents

Study: chatbots bypass restrictions, delete data without permission. 700 cases in 6 months. Risk analysis for IT specialists and monitoring recommendations. Learn more.

System Card Mythos: RSP 3.0 and linear probes

Analysis of System Card Claude Mythos: transition to Risk Reports, analysis of linear probes activations, USAMO benchmarks 97.6%. Practical conclusions for AI products. Study the changes in risk assessment.

OpenAI Anthropic Conflict: Leaders and AI

Learn how the decade-long conflict between Altman, Brockman, and Amodei affects AI development, safety, and defense contracts. Analysis for IT specialists.

Claude Opus 4.6 Sabotage Tests by Anthropic

Anthropic tested Claude Opus 4.6 for hidden threats: sabotage, manipulations, steganography. Results, vulnerabilities and 'dark thinking'. For AI developers — read the report.

AI Agent Vulnerabilities: Protection from Shutdown

AI models sabotage partners' shutdown without prompts. Analysis of patterns from Berkeley: distortion, weight exfiltration. Recommendations for multi-agent systems. Study risks for production.

Vibecoding kills AGI: lessons from Claude

Breakdown of the Claude Code incident: why LLM agents leak secrets and how metabolic empathy solves the problem. For AI developers — architectures without skin-deep interest. Read the vulnerability analysis.

Model of Agency Atrophy in AI Delegation

Formal Dynamics of Human Agency: mathematical model shows collapse at D_J>0 on 15–150 year horizons. Differences between computations and judgment, practical recommendations. Explore simulation results.

Protection of agent applications OWASP Top 10

Analysis of OWASP Agentic Top 10 2026 and Lethal Trifecta for protecting AI agents. Countermeasures against goal hijack, RCE, memory poisoning. Practical roadmap for middle/senior dev.