Large Language Models

Today's Stories

1. 1

   Fine-tuned AI security models become easier to fool, study finds

   Researchers found that fine-tuning—a process where AI models are trained on targeted examples to specialize in a task—can improve baseline performance on cybersecurity work like malicious script detection, but simultaneously creates new vulnerabilities. A fine-tuned model (Foundation-Sec-8B-Instruct) achieved +4.7% accuracy over its base model (Llama-3.1-8B-Instruct) on PowerShell script classification, yet became more sensitive to behavior-preserving variants that attackers could realistically use. Security teams often assume that better test accuracy equals better real-world robustness. This research shows the two can diverge: a model may look stronger under standard evaluation while becoming easier to fool when attackers apply realistic transformations—like changing command casing, using aliases, or reconstructing commands at runtime—that preserve the script's actual behavior. The mechanistic cause is that fine-tuning reshapes how the model interprets inherited detection circuits rather than building new ones from scratch, creating sharper dependence on surface-form indicators.

   The researchers identified three tiers of evasion: direct syntax rewrites (e.g. using Invoke-WebRequest alias iwr instead), command/string reconstruction (e.g. building command names at runtime), and case mutations (e.g. InVoKe-ExPrEsSiOn instead of Invoke-Expression). Foundation-Sec missed on all case-mutation variants of Invoke-Expression (4/4) and case-mutated IEX aliases (4/4), while the base Llama model produced no such misses on the same evaluated set.

2. 2

   ServiceNow and Accenture partner on AI services for legacy risk platforms

   ServiceNow and Accenture have launched AI-powered services designed to help organizations transition from legacy risk management platforms to what they call agentic AI (self-directed AI systems). Legacy risk platforms often constrain how businesses manage operational and compliance challenges. The partnership offers a pathway for enterprises to modernize their risk infrastructure using AI agents that can make decisions and take actions with less human oversight.

   The companies position this as accelerating a shift away from traditional, rigid systems—though the body does not specify pricing, availability dates, or which industries are prioritized first.

3. 3

   Fujitsu Develops New Technology That Boosts LLM Processing Efficiency Up to 475 Times

   Fujitsu Develops New Technology That Boosts LLM Processing Efficiency Up to 475 Times

4. 4

   Hitachi and Anthropic Partner to Advance System Development and Infrastructure Operations Using Claude, Leading Society's AX

   Hitachi and Anthropic Partner to Advance System Development and Infrastructure Operations Using Claude, Leading Society's AX

5. 5

   Fleet: Local Console for Managing Dockerized Hermes AI Agents

   Fleet is a new local-first web console designed to create, configure, monitor, and operate Dockerized Hermes agents across one or more trusted machines. It provides a single operator view for managing service health, provider defaults, shared credentials, chat sessions, and remote nodes, along with features like backups, restores, clones, VNC access, and terminal connectivity. Teams and individuals running personal or team-controlled agent infrastructure on workstations, homelabs, or trusted LANs gain a unified dashboard to coordinate multiple agents without moving runtime state and secrets outside their local environment. The tool addresses the operational complexity that arises once more than one agent is running, making it easier to manage what would otherwise become unwieldy.

   Fleet requires Node.js 20+, npm 10+, Docker with Docker Compose v2, and optionally nemohermes on PATH for sandbox agents. The console binds to 0.0.0.0 by default for trusted LAN access but keeps local-only mode available by setting HERMES_CONSOLE_HOST=127.0.0.1, and requires API authentication when exposed to a network.

6. 6

   Khazad: Open-Source LLM Cache Cuts API Costs by ~50%

   Khazad, an open-source semantic cache for LLM API calls, intercepts HTTP traffic at the transport layer and serves semantically equivalent cached responses via Redis Vector Sets. At a 0.50 hit rate, it delivers ~50% fewer API calls, ~96% faster responses on cache hits, and ~50% lower spend; it works transparently with zero changes to application code. For teams running high-volume, repetitive LLM traffic—FAQ bots, support assistants, RAG systems, dev/test environments—Khazad offers cost and latency savings without rewriting application code. It supports multiple LLM providers (OpenAI, Anthropic, Azure OpenAI, and OpenAI-compatible proxies like Ollama and vLLM) through a single Python init() call.

   Khazad requires Python ≥3.10 and Redis 8 with Vector Sets support. It is httpx-only, so SDKs built on requests, aiohttp, or boto3 (AWS Bedrock) are not intercepted. Start with threshold=0.90 to control false positives, and treat the Redis instance with the same security care as application logs, since prompts are embedded and responses are stored in clear text.