Today’s lead · thehackernews.com
HalluSquatting can steer AI coding assistants into installing attacker-controlled packages
New research describes an attack that registers plausible package or plugin names that AI assistants hallucinate, then uses prompt injection in fetched content to get attacker code run on a user machine. The researchers frame the pattern as a path to botnet-scale compromise when agents can fetch, install, and execute with little review.
Top signals
6 more
OpenAI releases GPT-Live for full-duplex ChatGPT Voice
OpenAI released GPT-Live-1 and GPT-Live-1 mini, voice models that can listen and speak at the same time for interruptions, silence handling, and live translation. GPT-Live-1 mini replaces the current Advanced Voice Mode by default, while paid users can access the larger model.
SpaceXAI releases Grok 4.5 with Cursor-trained coding and agent focus
SpaceXAI released Grok 4.5, positioning it as a token-efficient model for coding, app-building, research, writing, and routine knowledge work. Cursor says the model was trained jointly with SpaceXAI on developer-agent interaction data and ships with additional cybersecurity safeguards.
AI coding agents are triggering endpoint security rules built for intruders
Sophos telemetry found Claude Code, Cursor, and OpenAI Codex activity setting off behavioral detections for credential access and execution, including DPAPI browser credential reads, Windows credential-store enumeration, system-tool downloads, and startup-folder writes. The agents were not malicious, but their normal work resembled attacker tradecraft.
AMD, OpenAI, and Microsoft contribute MRC AI networking work to OCP
AMD said it is contributing MRC to the Open Compute Project with OpenAI and Microsoft, targeting scalability, resilience, and performance for large AI clusters. The company says the work is already deployed at scale in test clusters.
GitHub ‘Verified’ commits can be replayed under new hashes without breaking signatures
New research shows a signed Git commit can be rewritten into a different hash while preserving identical files, author, date, and GitHub’s “Verified” status. The issue does not let attackers change signed code, but it weakens blocklists, provenance logs, and deduplication systems that key only on verified commit hashes.
Anthropic and AE Studio propose GRAM modules as an off switch for dual-use model knowledge
Anthropic and AE Studio described GRAM, a preliminary training method that routes dual-use knowledge such as virology or cybersecurity into removable Transformer modules. Anthropic says GRAM has not been applied to production models and may never be.
Tools & repos
6 selected
Codex CLI 0.143.0
Codex CLI 0.143.0 turns remote plugins on by default, adds npm marketplace sources, supports macOS and Windows system proxies including PAC/WPAD, adds remote-control pairing, and improves MCP tool search. The release also adds Bedrock Sol, Terra, and Luna models with max reasoning support.
TencentDB Agent Memory
TencentDB Agent Memory provides fully local long-term memory for AI agents through a four-tier progressive pipeline with no external API dependency. It is relevant for teams experimenting with agent memory while keeping retrieval and storage inside their own environment.
ZML/LLMD
ZML released LLMD, a free inference server for running open-source LLMs across Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc hardware. The pitch is fewer chip silos and more flexibility for teams optimizing inference cost and supply.
Glean AI Gateway
Glean AI Gateway is a platform layer for routing across 30+ proprietary and open-source models, including OpenAI, Claude, Gemini, and NVIDIA, while centralizing security, usage visibility, and cost controls. It is aimed at enterprises trying to decouple fast-changing AI interfaces from shared governance and model-routing infrastructure.
ExploreYC
ExploreYC is an open-source API for querying Y Combinator and a16z company data. It is useful raw material for founder research, investor sourcing, market maps, and internal scouting tools.
last30days-skill
last30days-skill is an AI agent skill for researching a topic across Reddit, X, YouTube, Hacker News, Polymarket, and the web, then synthesizing a grounded summary. It is a useful pattern for packaging repeatable research workflows as agent skills.
Blogs worth your time
7 reads
Tuning the harness, not the model: a Nemotron 3 Ultra playbook
LangChain walks through how it raised Nemotron 3 Ultra’s Deep Agents performance by changing prompts, tool descriptions, and middleware rather than model weights. The useful part is the eval-driven harness loop: screen cheaply, inspect traces, make one targeted change, re-run, and keep only fixes that generalize.
Native-speed vLLM transformers modeling backend
Hugging Face says the Transformers modeling backend in vLLM is now at or above native vLLM speed for many architectures. This reduces the porting tax for new model authors: implement cleanly in Transformers and get fast vLLM serving without waiting for a bespoke backend.
Vercel Agent: An agent you can let near production
Vercel describes its dashboard, GitHub, and CLI agent for investigating production incidents, reviewing PRs, and proposing fixes under its own identity. The practical signal is the control model: read-only by default, approval-gated actions, and production context from logs, metrics, and deployments.
Automating cross-repo documentation with GitHub Agentic Workflows
GitHub’s Aspire team explains how merged product changes trigger agent-generated documentation PRs in another repo, with subject-matter-expert review before merge. The case is concrete: 82 feature-docs PRs for Aspire 13.3 and 13.4 merged with a 44.8-hour median lag after the product PR.
Flint: A visualization language for the AI era
Microsoft Research introduces Flint, an open-source chart specification language designed for AI agents to generate polished, human-editable visualizations. It uses semantic types, adaptive layout, multiple render backends, and an MCP server so agents can validate and render charts inside workflows.
Data for Agents
NVIDIA argues that real agent capability depends on data for tool-use failures, recovery, multi-step workflows, retrieval, safety, and user simulation—not just model weights. The post is useful for teams building eval and training pipelines because it frames agent improvement as an open-data and synthetic-data scaling problem.
Separating signal from noise in coding evaluations
OpenAI analyzes issues in SWE-Bench Pro and warns that popular coding benchmarks can give unreliable signals about model quality. For teams buying or routing coding agents, the takeaway is to treat public benchmark deltas as inputs, not deployment evidence.
Funding & acquisitions
7 moves
Prime Intellect
Prime Intellect raised a $130M Series A at a $1B valuation to expand its full stack for enterprise agent development. The platform combines compute access, reinforcement-learning tooling, and evaluation infrastructure so companies can train agentic systems without relying entirely on frontier labs.
SambaNova Systems
AI chip company SambaNova raised $1B at an $11B valuation in the first close of its Series F. The financing follows its SN50 chip launch and deepening Intel partnership for AI inference development.
Adage Automation
Goa-based Adage Automation raised ₹230 Cr to expand AI-enabled industrial software, advanced engineering, digital monitoring platforms, manufacturing, and global operations. The company builds gas analyser and environmental monitoring systems for process industries including oil and gas, steel, cement, chemicals, power, and energy.
Elevate Education
Elevate Education, formerly Sunstone, raised ₹170 Cr in Series D funding to strengthen AI and technology capabilities and expand its partner institution network. The higher-education platform supports over 25,000 students and expects ₹300 Cr revenue with profitability in FY27.
Milo Drive
Milo Drive raised $2.4M in seed funding for its EV mobility platform for cab drivers and small fleet operators. The startup uses data analytics and machine learning for ride allocation, charging guidance, fleet tools, and driver earnings transparency.
eNLife Research
Bengaluru-based eNLife Research raised ₹6 Cr to build and validate a blood-based biomarker platform for early detection of Alzheimer’s and other neurodegenerative diseases. The startup combines biomarker detection with machine learning and is developing India-specific datasets with research institutions.
Edysor.ai
Udaipur-based Edysor.ai raised ₹1.2 Cr for its voice AI platform that calls, qualifies, follows up with, and converts enterprise leads. The startup says it added 15 enterprise customers in two months and prices on outcomes rather than runtime.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.