Lakera Bulletin - This Week in AI #39: Wrapping Up 2025, Looking Ahead

[_Val_]

As the year winds down, AI isn’t slowing, it’s sharpening. This issue captures some of the latest developments heading into the end-of-year break, from concrete progress on AI security and governance to new model releases that hint at where the ecosystem is headed in 2026.

Let’s jump right in.

In case you missed it: 2025, the Year of the Agent

Lakera’s latest research analyzes real-world attacks observed across agent-enabled systems in Q4 2025, revealing how quickly attackers adapted to tool use, browsing, and multi-step workflows. The findings show indirect prompt injection, system-prompt leakage, and external content manipulation emerging as dominant risks heading into 2026.
🔗 Read the full analysis

UK AI Security Institute publishes first Frontier AI Trends Report

The UK’s AI Security Institute released its inaugural public report detailing what frontier AI systems can do today, and where risks and safeguards are evolving. It offers a rare, evidence-based snapshot aimed at improving global transparency around advanced AI capabilities.
🔗 Read the official announcement

NIST adds AI to its cybersecurity framework guidance

NIST introduced a new draft profile mapping AI risks into the widely used Cybersecurity Framework. The guidance helps organizations systematically address AI-specific threats alongside traditional cyber defenses.
🔗 Read the coverage

Google rolls out Gemini 3 Flash as its new lightweight flagship model

Google introduced Gemini 3 Flash, a faster and more cost-efficient model now powering Search’s AI Mode and other products. The release highlights how performance, speed, and scale are becoming table stakes in consumer AI.
🔗 See the announcement

Nvidia unveils Nemotron 3 open-source models

Nvidia expanded its open-source lineup with Nemotron 3, starting with a Nano variant focused on efficiency and reasoning. The move reinforces the growing role of open models in the AI ecosystem.
🔗 Read the Reuters report

Meta works on next-generation AI models

Meta is developing new AI systems for text, image, and video generation, reportedly code-named Mango and Avocado. The effort signals Meta’s push to stay competitive in multimodal and coding-capable AI ahead of 2026.
🔗 Read more

OpenAI and Anthropic deploy new age-safety measures

ChatGPT and Claude began rolling out systems to detect and manage underage users automatically. The changes reflect rising expectations that safety controls be built directly into widely used AI products.
🔗 Read the coverage

From clearer guardrails to faster, more accessible models, the close of 2025 highlights a familiar tension: rapid capability gains paired with growing responsibility.

Thank you for reading along this year, we’ll be back in your inbox in early 2026.

Happy Holidays and Happy New Year! And see you in 2026!

[the_rock]

Truly enjoyed this one...happy holidays!