← AM·PM Brief

IBM 2 models vs hidden prompts; Cmpnd ships dspy-wsp | AI Daily Brief

· Morning brief · 8 news · 7:48

Audio in Mandarin Chinese · English transcript below

⚡ IBM models target hidden prompt injection; open-source LLM closes security gap.

The LLM technology stack is now seeing a wave of new tools that bypass conventional training and directly intervene in model behavior—a practical expression of emergent capabilities in the stack. Taken together, weight-space programming and external memory modules attached to frozen models reshape model capabilities along two dimensions: internal editing and external extension. Meanwhile, structured annotation that isolates malicious instructions marks the engineering community’s turn toward systematic input controls to compensate for the limits of model-only safety.

IBM 2 models vs hidden prompts; Cmpnd ships dspy-wsp | AI Daily Brief

Today's Top 3 Headlines

  1. AI Industry News

    🤖 IBM open-sources two small models, prompt injection attack success rate falls to 0.1%

    IBM’s 3.7B/8.8B Granite LLMON previews on Hugging Face tag-split instructions/docs, cutting attack success 96–100%→≤0.1%. Devs get local commercial prompt-injection defenses, but…

    Source ↗
  2. AI Industry News

    🤖 Cmpnd launches dspy-wsp: LLM weight-space programming, up to 20-point gains

    Cmpnd's dspy-wsp adds weight-space programming to DSPy, lifting 0.6B–9B models up to 20 pp. It allows gradient-free weight patching, but 78–167-min runs keep it experimental.

    Source ↗
  3. AI Industry News

    🤖 Stanford open-sources Memorilla: 16 tokens compress entire library, 27x faster long text

    Stanford open-sources 194M-param Memorilla on GitHub: full library to 16 tokens, FactKG 84.33%, 108x token savings. Frozen LLMs skip retraining; one H100: 512 docs 27x faster.

    Source ↗

+5 more headlines

  • 🛡️ AI-driven new ransomware surges; low-barrier attackers can also cause severe damage
  • 🤖 FAA: no cyberattack evidence; Cessna completes 3,200-mile no-touch controlled flight
  • 🤖 Chatbots hit a wall; DeepMind bets on world models for simulated learning
  • 🤖 SenseNova releases Loop-9B preview on Hugging Face
  • 🤖 New Home Assistant box: one-click connect old appliances to smart home
Unlock all 8 headlines + deep analysis →Free 7-day trial · cancel anytime
Browse all past briefings →