Research on LLM collaboration and multi-agent frameworks is expanding, but studies are testing whether communication gains are real or just fine-tuning. Security concerns range from conversational backdoors to fragile safety decisions and open-weight models stripped of cybersecurity capabilities.
Kimi K3.1 and Tencent Hy4 remove cybersecurity, reverse engineering, and patching capabilities from open-weight models. If local models cannot analyze vulnerabilities or binaries, auditing on your own hardware becomes impractical and teams fall back ...
PLatour shows that part of the gains in PACT systems appears even with non-informative senders or when only the receiver is trained. PACT-Check measures the real contribution of inter-model communication. For on-premise deployments, the difference be...
An open-source framework challenges the PACT premise: passing latent representations between models can improve benchmarks without real information transfer. PLatour adds trainable zero-initialized bridges and PACT-Check, a method to separate genuine...
An attack framework uses poisoned dialogue data to steer multi-turn interactions after a trigger, without modifying prompts at deployment time. Three variants — Direct, Constrained Progressive, Unconstrained Progressive — show that models preserve ap...
A single-model multi-agent framework shares one backbone and coordinates GPU-resident KV states through a directed acyclic graph. Zero-copy radix forking, attention-only KV mounting, and packed DAG-GRPO training cut peak KV memory by 2.8x and raise t...
A benchmark on 3,000 predictions from ten frozen open models shows 86.1% appear plausible, but authorization remains fragile: low ASK recall, strong BLOCK bias. SafeWebWorld and LAR separate evidence preservation from decision calibration. For local ...
