CP SignalTechnology worth your attention.

Monday, 12 October 2026

The Daily Edition: 4 stories, ranked by significance

  1. Artificial intelligence

    AI Models Can Provably Hide Arbitrary Capabilities

    A NeurIPS 2026 paper demonstrates that AI capabilities can be hidden inside model weights using encryption so that they only activate on a secret trigger, remaining undetectable even to someone with full white-box access to the model. The authors tested standard elicitation techniques — fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training — on a proof-of-concept chemical reaction prediction task (a stand-in for CBRN-relevant capability) and found the hidden capability resisted all of them, backed by cryptographic hardness guarantees rather than just empirical robustness.

    Read the storyNo. 1 of 4arXiv cs.CRFri 9 Oct
  2. Artificial intelligence

    ToFu: Exploiting and Fuzzing the LLM-Orchestrator-Tool Trust Boundary

    Researchers introduce ToFu, a framework that automatically derives security checks from MCP tool metadata and fuzzes tools to find cases where documented input semantics aren't actually enforced. Testing 273 open-source MCP servers (3,405 tools), they found verified bugs tied to known CWE weakness classes in 343 tools spanning 118 servers, showing that LLM-generated tool arguments are frequently trusted when they shouldn't be.

    No. 2 of 4arXiv cs.CRFri 9 Oct
  3. Artificial intelligence

    APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces

    A new benchmark, APEX-Harm, tests 12 frontier LLM agents against prompt and script injection attacks while they perform finance, consulting, and legal tasks in large simulated workspaces with MCP-exposed tools. Even top-performing models like Opus 5 and Sonnet 5 (8% and 20% attack success) remained highly vulnerable to task-specific script replacement and attacker-supplied generic scripts, which succeeded 73-84% of the time across models regardless of whether the agent completed its assigned task correctly.

    No. 3 of 4arXiv cs.CRFri 9 Oct
  4. Artificial intelligence

    Same Query, Different Answer: Run-to-Run Variance in Semantic Query Processing

    Researchers re-executed three semantic query benchmarks (SemBench, LRO-Bench, UDA-Bench) across 1.2 million model calls and found substantial run-to-run variance even at temperature 0, with up to 24% of rows changing output on one task. In SemBench's own published results, 21% of text results changed quality between executions, and most of LRO-Bench's implementation recommendations were not actually settled by the benchmark's executions.

    No. 4 of 4arXiv cs.CRFri 9 Oct