Today in AI

Enterprise AI deployments hit a credibility wall today, and the warning signs are coming from multiple directions. A sobering arXiv paper reveals that AI agents routinely strip away critical formatting cues—bold text, tables, indentation—when ingesting documents, then proceed to make "confident errors" in policy interpretation. This isn't a model-specific quirk; it affects both Claude and GPT models. The kicker? These systems often don't know what they've lost, which dovetails neatly with another sharp critique making the rounds: using LLMs for confidence scoring is fundamentally broken. The Hacker News analysis argues that LLM-generated confidence scores lack genuine probabilistic calibration—they're theater, not telemetry. When your enterprise agent silently mangles a compliance document and then tells you it's 97% sure of its wrong answer, you've got a compound failure that no amount of RAG marketing can paper over.

Meanwhile, the infrastructure beneath these systems is showing real strain. LiteLLM 1.107.0 on PyPI was compromised with a credential stealer and backdoor—a supply chain attack where the GitHub source looked clean but the distributed package was weaponized, triggering on any Python process startup. It's the kind of attack that makes "trust but verify" feel quaint. On the brighter side, OpenAI's Patch the Planet program used GPT-5.6 Sol to uncover critical PHP2Shell vulnerabilities affecting 70% of the web, coordinating disclosure with Google's Wiz over two months. And NVIDIA's Cosmos platform continues its push into physical AI with new tools for world model deployment—less headline-grabbing, but foundational for robotics and simulation. The research front stays active too: VX-One tackles honest, calibrated security triage under drift, while Quant-Arena benchmarks LLM agents on real trading data—both signs that domain-specific rigor is where the real evaluation battles are being fought.

**Bottom line:** AI's enterprise readiness gap is widening—silent document corruption and fake confidence scores are production risks today, while supply chain security and domain-specific validation remain the hard problems worth solving.