Frontier Model Watch

Benchmark Innovations

• 🔥 PostTrainBench: POSTTRAINBENCH benchmark evaluates LLM agents like Claude Code and Codex CLI on autonomously performing...

Critical security flaws in publicly accessible Ollama LLM servers:

• Unauthorized model access via exposed inference APIs
• Prompt abuse by attackers
• GPU resource consumption hijacking compute
  Secure your deployments now.

Emerging techniques highlight LLM vulnerabilities in a new wave of red-teaming:

• Activation steering like Amnesia triggers compromised, offensive...

Arena has become the de facto leaderboard for frontier LLMs, driving funding, launches, and PR—but faces integrity tests:

• Explosive growth: UC...

PostTrainBench benchmarks LLM agents like Claude Code automating post-training on Qwen/Gemma across math, coding, science.

• Progress: Outperform...

NAVER's Seoul World Model (SWM) pioneers promptable, grounded world sims in real Seoul:

• Multi-km trajectories without visual drift, via street-view...

Randy Goebel unveiled a partial framework for debugging foundation models at Ontology Summit 2026:

• LLM Flaws: Popular but produce incorrect...

Promptfoo showcases AI red teaming at RSA Conference 2026:

• Live demos of LLM attack vectors
• Jailbreak techniques in action
• Automated red teaming workflow integration
  Essential for frontier model security.

Open Source Security Foundation launches Model Signing v1.0 to secure the machine learning supply chain, building on SLSA frameworks from traditional software. A key open-source advancement tackling security gaps in accelerating ML ecosystems.

Open Model Releases & Partnerships

• 🔥 NVIDIA Expands Open Models: NVIDIA expanded its open model families targeting autonomous agents, physical...

Anthropic advances AI safety with a dedicated hire: a manager to design and implement protections against chemical and explosive threat risks. This underscores their commercial priority on preventing catastrophic misuse.

• Rapid LLM adoption in financial services introduces operational, regulatory, security risks
• Secure deployments demand layered defenses: system...

• API bills trigger self-hosting: Teams question running LLMs on-prem after first real bills arrive.
• Unified APIs across providers: LiteLLM proxies...

Key eval progress for DeepSeek-R1, Gemini 3.0, ChatGPT-5 equivalents:

• Reasoning advances (DeepSeek-R1 Pure RL, Gemini Deep Think) boost accuracy but...

LLMs are increasingly deployed as tool-using agents, shifting safety concerns from harmful text generation to harmful task completion.

Frontier labs eye neurotech: Sam Altman co-founded Merge Labs, securing massive early 2026 funding for ultrasonic brain-computer interfaces.
-...

Key US governance tensions expose frontier AI weaknesses:

• Narrow statutes target models over 10²⁶ FLOP, defining critical safety incidents.
  -...

A new international AI Safety Report argues frontier capabilities are advancing faster than mitigations, coinciding with model self-explanations coming under question from a cross-lab paper. Urgent call for faster safeguards.

OpenAI advances frontier deployment on two fronts:

• GPT-5.4 Thinking released as latest GPT-5 series reasoning model, with comprehensive safety...

Key advances in reasoning and agents from latest research:

• Phi-4-reasoning-vision-15B: Microsoft's multimodal reasoning model
• LLM agent memory:...