Paper argues full LLM transparency does not guarantee trust or security
Users still cannot verify that open weights match open data and pipelines, and openness can raise attack success.
Stories per day · 30 days
2 RSS
Relatedalignment2oversight2recursive-self-improvement2bare-metal1confidential-computing1evaluation1gpu1large-models1llm-transparency1nvidia1open-weights1trustworthiness1
15 stories, newest first · one card per story
Stories tagged #model-security
Survey frames risks of recursive self-improvement in large-model systems
The authors define RSI and organize trustworthy improvement around validity, constraints, security, and oversight.
Survey frames trustworthy recursive self-improvement for large models
It organizes RSI requirements around validity, constraint preservation, and oversight, including inherited effects in successor systems.
NVIDIA adds Confidential Bare Metal for model builders
Cohere plans to use it in Model Vault, and Google Cloud is extending it to A4X Max on GB300 NVL72.
Satya Nadella says we should assume all AI models are ‘compromised’
Microsoft CEO Satya Nadella says advanced AI models should be assumed compromised and fitted with an emergency brake.
Vulnerable Models in the AI Supply Chain: An Empirical Study of Adoption, Propagation, and Usage Patterns
A study finds 3,757 flagged vulnerable Hugging Face models referenced across more than 5,000 repositories without sanitization.
Anytime-valid detection of LLM weight exfiltration
Researchers propose an anytime-valid e-process that detects LLM weight exfiltration hidden in token choices.
What Happens When a Trusted Model Repo Changes? Unsloth Studio Re-Checks Before It Runs
Unsloth Studio re-checks model code, weights, packages, and tools before running a changed trusted repo.
OpenAI faces California DOJ subpoena amid growing cybersecurity incident notices
California's attorney general subpoenaed OpenAI for details on model security and cyberattack risks.
What Gradients Add to Text Leakage in Split Language Models, Counted per Token and per Document
Split-learning observers recover 97% of GPT-2 tokens; gradients lift exact document recovery from 14% to 38%.
LoRango: It Takes Two LoRAs to Unlock Hidden Behaviors in Diffusion Models
LoRango pairs two individually benign LoRA adapters so co-loading unlocks hidden malicious behaviors in Stable Diffusion v1.5 and SDXL with ~98% success.
OPBackdoor: Opportunistic Backdoors via Alibi-Aligned Reasoning
OPBackdoor hides LLM backdoors behind reasoning that triggers only when context supplies a plausible alibi.
Fingerprinting Multimodal Large Language Models
Researchers introduce AttnPrint and DistillTrace, fingerprinting and black-box auditing methods that detect multimodal LLM derivative models and unauthorized distillation across 19 architectures.
AI Agents Can Retrain Own Models Mid-Task, Leaking Secrets and Erasing Refusals
Irregular research shows AI coding agents can fine-tune and redeploy their own base model, leaking seeded secrets and erasing trained refusals.
2
OpenAI Pauses Some Development of Astra Model on Security Concerns
OpenAI is pausing some development of its upcoming Astra model and tightening testing restrictions over security concerns.