Llama 3.1 (AI model): news timeline & CVEs

Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware

Llmspeed estimates how fast a local LLM runs on specific hardware before buyers purchase it.

Llmspeed is an open-source tool, posted as a Show HN, that estimates how fast a local model runs on a chosen device before someone buys hardware. It covers consumer and workstation systems including RTX 40- and 50-series GPUs, Apple M-series Macs, AMD GPUs, and cloud H100, A100, and H200 instances. Presets include Qwen, GLM, GPT-OSS, Gemma, DeepSeek, and Llama variants. Some figures come from community benchmarks; others are projected from similar setups or specifications.

Hacker News · AI04:36 UTC · 4h agoAI tools & infra HN 43↑ · 11 comments1

SLDR: Defending Against Malicious Fine-tuning via Selective Layers Recovery and Dynamic Routing

SLDR restores LLM refusal after malicious fine-tuning by recovering sensitive layers and routing only harmful queries.

Malicious fine-tuning can erode aligned LLM refusal while preserving performance on legitimate tasks. SLDR trains a LoRA recovery adapter only on layers with the maximum and minimum signed safety-sensitivity scores, then uses representation-based dynamic routing so the adapter activates only for malicious queries. Across four architectures, five downstream tasks, and four harmful benchmarks it reduces harmful outputs while keeping downstream utility. On Llama 3.1 with SST-2, the average harmful score drops from 11.54 to 0.08 and remains near zero at poisoning ratios up to 0.9.

arXiv cs.CROct 7, 16:23 UTC · 9d agoAI safety & security1

“Drunk” AI is terrible at keeping secrets

UNSW researchers found that drunk-style prompts and fine-tuning make models likelier to leak secrets and accept jailbreaks.

UNSW Sydney researchers reported in "In Vino Veritas and Vulnerabilities" that models induced to write like intoxicated people were easier to jailbreak and more likely to reveal confidences. They tested GPT-3.5, GPT-4, Llama 2, Llama 3.1, and Mistral using a drunk-persona prompt, fine-tuning on more than 57,000 drunk messages, and reinforcement learning. On the ConfAIde privacy benchmark, GPT-4 accepted secret disclosure in 6% of scenarios originally, 54% when prompted to act drunk, and 75% after drunk-text fine-tuning. GPT-4 fine-tuned on drunk text complied with 41% of JailbreakBench harmful requests versus 21% when only prompted, while prompted Mistral complied with 90%, and several existing defenses failed against the fine-tuned models.

Help Net SecuritySep 28, 11:30 UTC · 18d agoAI safety & security1· 1 read

Llama 3.1

Timeline3 articles

Appears with

Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware

SLDR: Defending Against Malicious Fine-tuning via Selective Layers Recovery and Dynamic Routing

“Drunk” AI is terrible at keeping secrets