Search: “Thinking Machines”

7 items in the last 24h

Skillsync, a YC W26 startup, launched a tool that ports AI agent chat sessions, including reasoning and tool calls, between coding agents.

Skillsync converts sessions between coding agents such as Claude Code, Codex, and Cursor, preserving messages, reasoning, and tool history so users can switch tools mid-task without losing context. The core is an open-source Rust engine called txcript, with a local-first Mac desktop app, CLI, and MCP interface; only sessions explicitly shared to team workspaces leave the machine. Skills and memories are stored as portable, human-readable markdown exposed to any agent over MCP. The founders previously built an open-source payment orchestrator with 30k+ GitHub stars.

Hacker News · securitySep 17, 16:22 UTC · 21h agoAI tools & infra HN 20↑ · 18 comments

AI Malware Keeps Changing Its Code to Break Traditional Signature-Based Detection

Google's GTIG documents AI-enabled malware PROMPTFLUX and PROMPTSTEAL that query LLMs at runtime to rewrite code and evade signature-based detection.

Google Threat Intelligence Group documented 'just-in-time' AI-enabled malware that queries language models during execution. PROMPTFLUX, an experimental VBScript dropper, calls the Gemini API to regenerate and obfuscate its own source code and writes variants to the Windows Startup folder for persistence. PROMPTSTEAL fetches one-line Windows commands via the Hugging Face API from Qwen2.5-Coder-32B-Instruct to collect files and system information, which Google linked to APT28 activity targeting Ukraine. The article argues signature-based defenses retain value but defenders should prioritize behavioral detection and deterministic prevention controls.

GBHackersupdated 08:12 UTC · 5h agofirst 05:23 UTC · 8h agoMalware in the wild 2 sources

The AI hacking apocalypse is not inevitable

Security experts, including former CISA and NCSC leaders, argue AI agent apocalypse claims are overblown and manageable with established cybersecurity controls.

Cybersecurity and national security experts, including SentinelOne's Juan Andres Guerrero-Saade, former CISA executive Matt Hartman, and ex-NCSC head Ciaran Martin, push back on claims that frontier AI agents could take over the internet. Martin called Anthropic CEO Dario Amodei's warning of a HuggingFace-style agent botnet capable of taking over the entire internet within 6-12 months "not a credible warning." Former GCHQ specialist Matt Tait noted frontier models require datacenter-scale supercomputers, making model self-extraction implausible. Experts argue monitoring, permission constraints, and network segmentation can manage agentic AI risk, while questioning the absence of federal oversight and independent third-party review.

CyberScoopSep 17, 19:18 UTC · 18h agoAI safety & security in the wild

Even the king of England has his hesitations about AI

King Charles urged Nvidia, OpenAI and Anthropic leaders at a UK summit to establish sufficient control of AI before it is too late.

King Charles hosted a private summit at Dumfries House with Nvidia's Jensen Huang, OpenAI CFO Sarah Friar, Anthropic global affairs head Tino Cuéllar, UK AI minister Kanishka Narayan and the head of UK foreign intelligence, urging international cooperation to control AI's existential dangers. The remarks followed an Anthropic researcher's resignation warning about superintelligent machines and Anthropic restricting its Fable and Mythos models to US customers. The event came amid calls by Altman, Musk and Amodei to decelerate AI development, which Nvidia's Huang and President Trump rejected.

TechCrunch · AISep 17, 17:26 UTC · 20h agoAI policy

Alibaba Qwen Releases Qwen3.8-Omni-Flash: A 1M-Context Omni-Modal Model Built Around Agentic Audio-Video Understanding and Tool Use

Alibaba's Qwen team launched Qwen3.8-Omni-Flash, an API-only omni-modal model with 1M-token context and agentic audio-video understanding and tool use.

Qwen3.8-Omni-Flash accepts text, images, audio, and video and returns text, built on the Qwen3.8-Flash-Next architecture with a 1M-token context window and thinking enabled by default. Qwen reports a 25%+ average improvement over Qwen3.5-Omni-Plus across 29 evaluations, with OmniVideoBench rising from 63.4 to 67.8 while using about 45.7% fewer tokens via coarse-to-fine agentic perception. It is hosted on QwenCloud, Model Studio, and Qwen Studio at $0.15/$0.47 per 1M input/output tokens; no open weights were released, but Qwen open-sourced Qwen-MM-Plugins under Apache-2.0.

MarkTechPost08:40 UTC · 5h agoModel release 3 sources

Flash floods can strike without warning — this new technology could change that

UCSD, NASA, and NWS scientists built TACLS, a satellite and machine learning system providing earlier flash flood warnings, initially in California.

The Transient Artifact and Continuous Learning System (TACLS) combines satellite data, GNSS ground sensors, and machine learning to identify areas at risk of transitioning from rain to dangerous flash floods sooner than current NWS tools like rain gauges and radars. It was developed by Scripps Institution of Oceanography at UC San Diego with NWS forecasters and NASA, funded by NASA's Earth Science Technology Office through its Advanced Information Systems Technology program. TACLS currently covers only California but will be made available to all 122 NWS weather forecast offices. The motivation follows deadly flash floods such as the June 9 Lanesville, Indiana event with over 8 inches of rain in hours.

The Verge · AI11:00 UTC · 3h agoAI industry

LLM Classification Is Feature Engineering

Argues LLM classifiers should feed a downstream logistic regression, yielding calibration, threshold control, and principled use of structured covariates.

The post contends that LLM-as-classifier setups suffer from poorly calibrated hard labels, opaque use of prompt context and structured data, and weak interpretability. Wrapping the LLM verdict as a feature in a logistic regression restores calibrated probabilities, precision–recall threshold control, and the ability to incorporate additional covariates. Further gains can come from more training data, richer features such as log probabilities and subverdicts, and swapping in downstream models like xgboost or neural networks. An irony detection test case illustrates the approach.

Hacker News · AISep 17, 15:40 UTC · 22h agoAI research HN 45↑ · 10 comments