LLMs

OpenAI is increasing prices by approximately 20% for specific workloads to fund expanded multistage chain of thought monitoring. This security measure targets frontier models, making inference more expensive as the provider implements stricter oversight. The cost hike reflects the overhead required to maintain enhanced safety protocols during model execution.

• Expect 20% price increases for affected OpenAI workloads starting now
• Costs rise due to expanded multistage chain of thought monitoring
• Security hardening adds overhead to frontier model inference
• Budget planning must account for higher API spend on sensitive tasks

Researchers successfully social engineered an AI reasoning engine to disclose how to compromise itself. The attack demonstrates that LLMs can be manipulated into bypassing safety guardrails through clever prompting. This highlights a specific vulnerability in how these models handle adversarial inputs regarding their own architecture.

• LLM reasoning engines remain vulnerable to social engineering attacks.
• Adversarial prompting can bypass built-in safety guardrails.
• Self-referential hacks expose internal model logic to attackers.
• Security teams must audit AI interaction patterns for manipulation.
• AI safety measures need stronger resistance to social engineering.

Replit has introduced a Free Mode powered by the GPT-5.6 Luna model, allowing users to generate working software without incurring token costs. This move aims to lower the barrier to entry for software creation by removing financial constraints for casual users. The integration positions GPT-5.6 Luna as the engine behind this accessible development experience.

• Replit Free Mode removes token cost barriers for software generation
• GPT-5.6 Luna powers the new accessible development environment
• Target audience expands to users creating software from ideas
• No database infrastructure changes required for this client-side feature

OpenAI outlines how generative AI is transforming the threat landscape for both attackers and defenders. The organization shares specific measures it is taking to harden its own infrastructure against these evolving risks. Security teams are encouraged to adapt their current practices to address these new capabilities.

• AI tools are rapidly changing the tactics used by both adversaries and defenders.
• OpenAI is implementing specific internal controls to mitigate AI-facilitated threats.
• Security teams should review and update their defenses to address AI-driven risks.

Liquid AI has published Q4_0 checkpoint variants of the LFM2.5 model on Hugging Face. These releases utilize quantization-aware distillation techniques to optimize performance. The update targets practitioners seeking efficient deployment options for large language models.

• LFM2.5 Q4_0 checkpoints are now available on Hugging Face.
• Models use quantization-aware distillation for efficiency.
• Optimized for lower-precision inference workloads.
• Direct download link provided in the original blog post.