Tag: AI Models
-
LLM Watermarking: How Models Embed Detection Signals in Their Outputs
Kirchenbauer’s green-red token list (ICML 2023), Aaronson’s EMS, and Kuditipudi’s ITS scheme all embed detectable statistical signals into LLM outputs. But Zhang et al. proved no watermark is…
-
Differential Privacy for LLMs: The Training Privacy Guarantee
Differential privacy provides the only formal guarantee against LLM training data leakage. DP-SGD’s four steps, the Moments Accountant, Renyi DP, and the epsilon values that actually mean something…
-
Multiagent LLM Security: When Your Agent Talks to a Malicious Agent
When LLMs call other LLMs as tools, injection attacks jump the boundary. ConVerse (2026) found 88% privacy violations and 60% security breaches in plausible agent-to-agent discourse. Here is…
-
LLMail-Inject: What 208K Attacks Against an Email Agent Found
Microsoft Research’s LLMail-Inject challenge: 839 participants, 208,095 unique attacks against a simulated email agent with production defenses. The finding: adaptive attackers breach even well-designed defense stacks. Here is…
-
Adversarial Machine Learning: From Szegedy to LLM Attacks
Szegedy (2014) showed deep networks could be fooled by imperceptible perturbations. FGSM, PGD, and C&W followed. By 2025, the same mathematical framework governs jailbreaks, poisoning, and memorization extraction.…
-
How RLHF and Constitutional AI Build Safety Into Language Models
RLHF trains models to prefer human-preferred outputs. Constitutional AI uses AI self-critique guided by principles. Neither provides formal guarantees. Here is how both techniques work, what they install…
-
LLM Training Data Memorization: When Models Leak Their Training Sets
LLMs memorize verbatim sequences from training data. Carlini et al. demonstrated extraction of phone numbers, email addresses, and private keys from GPT-2. Here is the mechanism, what gets…
-
LLM Supply Chain Attacks: PoisonGPT to Poisoned Skills
PoisonGPT used $1 of compute to pass benchmarks with modified facts. The April 2026 PoisonedSkills paper tested the same supply chain logic against Claude Code and Gemini CLI.…
-
Jailbreaking vs Prompt Injection: Two Different LLM Problems
Jailbreaking targets model content policy. Prompt injection targets application architecture. The defenses don’t overlap, the responsible parties differ, and the same RLHF training that resists jailbreaks amplifies injection…
-
LLM Excessive Agency: Why Every Tool Your Agent Has Is a Risk
Every tool an LLM agent has is an attack surface. OWASP’s LLM06 and the b3 benchmark across 31 models show why: capability scope determines blast radius. Here is…
-
Julia Bazinska and the Science of Measurable AI Security
Julia Bazinska built the empirical tools that make LLM security measurable. From DeepMind RL to first-authoring b3, here is what her research at Lakera actually produced.
-
Gandalf the Red: What 279K Real Attacks Reveal About LLM Defense
Lakera’s ICML 2025 paper ran 279K crowdsourced attacks to show what synthetic red-teaming misses. The D-SEC finding: system prompts degrade user experience without blocking attackers. Here is the…
-
Vision-Language Models: Architecture and the Benchmark Gap
How CLIP, SigLIP, Q-Former, and MLP adapters work in vision-language models. Why Qwen2.5-VL compresses visual tokens 4x, and what current VLMs still cannot do.
-
Chinchilla Scaling Laws: Three Methods and Why Labs Ignore Them
Chinchilla proved GPT-3 was undertrained. The 20:1 rule is a training-compute floor. Three methods, their disagreements, and why frontier labs now exceed it.
-
Speculative Decoding: How LLMs Generate 3x Faster
Speculative decoding achieves 3-4x LLM speedup with zero output quality loss. The math proof, EAGLE-2’s 4.26x result, and when it does not help.
-
LLMs in Veterinary Clinical Practice: What the Evidence Actually Shows
ChatGPT-4.5 scored 90% on feline eye disease cases vs 96.7% for experienced veterinary ophthalmologists and significantly outperformed novices (56-67%). Where LLMs add clinical value in veterinary practice, where…
-
Generative AI for Small Molecule Drug Discovery: How It Works and What the Evidence Shows
Generative AI is producing novel molecules from VAEs, GANs, and diffusion models. Machine learning virtual screening shows 75% hit validation rates against 106M-compound libraries. Why no AI-designed drug…
-
AI in Digital Pathology: What Computational Pathology Can and Cannot See
An NIH multi-institution study in Lancet Oncology classified 52 CNS tumor types from tissue images at 80% accuracy across 5,516 test samples. A Cancer Science paper simultaneously documented…
-
FDA Clearance for AI Medical Devices: What 510(k), De Novo, and PMA Actually Mean
The FDA has cleared 700+ AI medical devices through 510(k), De Novo, and PMA pathways. A March 2026 European Radiology review documents how the EU AI Act, FDA…
-
AI-Driven ADMET Prediction: What the Blind Challenge Results Actually Show
Deep learning beat classical methods for ADME prediction in a 65-team blind challenge at the 2025 OpenADMET competition. An AI-PBPK platform predicted full human pharmacokinetic curves from molecular…
You must be logged in to post a comment.