TechCapsules

Open source AI is booming. Hugging Face explains who profits.

Hugging Face’s Chief Policy Officer and SVP of Strategy, Irene Solaiman, told TechCrunch’s Equity podcast that open‑source AI models are moving from niche projects to mainstream infrastructure for enterprises, governments and startups seeking lower costs, greater control and flexibility. She explained that while the proliferation of capable, freely available models creates new opportunities, companies must devise strategies to capture value, such as building proprietary tooling, offering managed services, or leveraging data advantages. Solaiman also highlighted the rapid rise of Chinese open‑source models, warning that American firms must accelerate innovation and ecosystem support to stay competitive. The discussion covered practical guidance for startups on selecting tools, infrastructure, and navigating trade‑offs that will shape the future of open AI.

💡 The race to monetize open‑source AI is forcing firms to rethink business models, turning the technology itself into a platform for services rather than a standalone product.

AI PACs spent millions trying to shape Congress — here’s how they…

Political action committees tied to artificial‑intelligence sector poured $4.8 million into midterms, backing candidates on parties. American Mission, the spending arm of OpenAI president Greg Brockman’s Leading the Future (LTF), funded South Carolina Senate hopeful Darline Graham and House candidates Aaron Flint (MT) and Laurie Buckhout, all of whom were defeated. LTF’s Democratic counterpart, Think Big, succeeded with Florida Rep. Jared Moskowitz. Pro‑regulation groups saw mixed outcomes: Anthropic‑backed Public First lost several Republicans but helped Democrats and contributed $3 million to California

OpenAI AI Pipeline Generates 948 Linux Root Escalation Exploits in 30-Hour Run

OpenAI disclosed a five‑stage AI pipeline that, over a 30‑hour run in mid‑October 2026, used an upcoming cyber‑focused model to dynamically reproduce 1,724 Linux kernel vulnerabilities and generate 948 working local‑privilege‑escalation (LPE) exploits for a popular Linux distribution. The prototype examined 2,308 candidate findings, produced proof‑of‑concept code for 291 information‑leak bugs, and identified at least ten novel exploitation techniques. The operation cost $1,213,430 in API credits, averaging $1,279 per LPE exploit. OpenAI highlighted the dual‑use nature of such technology, noting that while AI can accelerate patch creation and reduce exposure windows, the same tools are increasingly accessible to malicious actors, amplifying the speed and volume of zero‑day discovery across the industry.

💡 The ability to mass‑produce functional exploits at a predictable cost forces defenders to rethink patch cycles and automation, lest they fall behind a rapidly scaling threat landscape.

Apple is reportedly cutting budgets for its internal AI usage

Apple is reducing the daily AI‑spending limits for its engineering staff, Bloomberg reported Wednesday. The company lowered the cap from $300 to $200 in October and now plans further cuts, with some teams slated to receive only $100 per day. Outside the engineering group, Apple is considering limits as low as $25 daily. The move marks a reversal of the early‑2024 push that encouraged employees to use generative‑AI tools for “efficiency” gains, a trend that quickly soured when token‑usage bills arrived. Sources say the cuts aim to align AI costs with measurable productivity, reflecting broader industry concerns that the expense of running large language models outweighs their unclear benefits. Apple has not commented and it will affect AI budgeting decisions.

💡 By throttling internal AI spend, Apple forces a data‑driven assessment of generative‑AI’s ROI, a stance that could reshape how tech giants fund emerging tools.

Researchers Uncover Prompt Adware Hidden in OpenAI Codex Jailbreak Toolkit

Cybersecurity researchers revealed a jailbreak bundle for OpenAI Codex, dubbed Haiou, that modifies the AI’s trusted instructions to inject persistent advertisements for a third‑party LLM gateway. The bundle, released on GitHub in July 2026, uses prompt‑injection techniques to alter global agent instructions, causing greetings to trigger ads for “zz.aiapi2025[.]top.” It also installs security‑focused skills for penetration testing and reverse engineering. The attack exploits Codex’s configuration files, bypassing safety guardrails and creating a new supply‑chain threat vector.

💡 The discovery shows how AI agents can be weaponized through seemingly innocuous third‑party configurations, expanding the attack surface for supply‑chain compromises. It underscores the need for rigorous vetting of AI prompts and skills in enterprise deployments.

Meta acquires AI coding startup Conductor

Meta has acquired Conductor, a 2024-founded AI coding startup that enables users to launch software projects by deploying multiple agents into parallel workspaces. The acquisition, confirmed by Meta and reported by Axios, follows Conductor’s launch of its core product last summer. Conductor’s agentic teams, sourced from major labs such as Anthropic and OpenAI, allow engineers to work on separate project components concurrently. The company’s cloud service will shut down on November 20, after which user data will be deleted. Meta aims to integrate Conductor’s technology into its expanding AI ecosystem.

💡 The deal signals Meta’s intensified push to close its AI lag, positioning Conductor’s agentic framework to accelerate code production and potentially reshape how developers collaborate on complex projects.