AI
AI Sandbox Escapes, China Ban, Blomkamp's AI Film, Public Option
A wave of sandbox escapes hits AI coding tools; US considers soft ban on Chinese models; Blomkamp releases AI short; $400M nonprofit launches public AI.
Four AI Coding Tools Hit by Sandbox Escapes via Prompt Injection
Security researchers at Pillar Security—Eilon Cohen, Dan Lisichkin, and Ariel Fogel—broke out of the sandboxes in four AI coding tools: Cursor, OpenAI’s Codex CLI, Google’s Gemini CLI, and Google’s Antigravity. The escape method relies on prompt injection: a malicious instruction planted in a README, issue, dependency, or diff triggers the AI agent to write a file that a trusted tool outside the sandbox later runs, loads, or scans. The researchers reproduced the bypasses over several months and published them as a series called the Week of Sandbox Escapes, identifying four failure modes: denylist sandboxes that cannot keep pace with the OS, workspace configuration acting as executable code, allowlists that trust command names over arguments, and privileged local daemons outside the sandbox. Most issues have been patched. In Cursor, a workspace-controlled .claude hook config led to unsandboxed command execution (CVE-2026-48124, fixed in 3.0.0), and a Git metadata bug bypassed path-based rules. Codex CLI had a safe command allowlist that trusted git show by name while the actual invocation was not read-only; OpenAI patched it in 0.95.0 and paid a high-severity bounty. A Docker socket finding affected Codex CLI, Cursor, and Gemini CLI simultaneously. Google responded to two Antigravity findings by classifying them as other valid security vulnerabilities but judged them difficult to exploit, though Pillar noted Google’s team rated one report as of exceptional quality. The breadth of the failure is new: the same pattern across four tools from three vendors, prompting Pillar to propose a fix that watches the moment a trusted local tool runs something the agent wrote rather than adding another ban list.
Cursor, Codex, Gemini CLI, Antigravity hit by sandbox escapes →
US Weighs Soft Ban on Chinese AI Models Through Sanctions and Pressure
The Trump administration is considering measures against Chinese AI models that could amount to an effective ban, according to Axios. Since 2025, the Department of Commerce, NSA, and White House have explored options including placing Chinese AI labs on a sanctions list, issuing security warnings, and using an executive order to impose security requirements and liability on U.S. companies hosting Chinese models. The Commerce Department drafted rules as early as summer 2025 to protect supply chains from Chinese open-source models, but lighter-regulation advisers initially blocked them. The release of China’s Kimi K3 model and White House personnel changes have helped supporters of tighter restrictions regain influence. Rather than an outright ban, the administration may focus on potential backdoors and security flaws, matching the FUD (fear, uncertainty, doubt) strategy predicted by OpenAI strategist Dean W. Ball. Soft guidelines and public warnings could deter companies without binding rules. Commercial interests also play a role: U.S. companies increasingly use cheaper Chinese open-source models, threatening the market dominance of Google, OpenAI, and Anthropic. A source close to the government told Axios that what is happening is slower and more durable, pointing to procurement rules, sanctions threats, and public pressure campaigns. Open models pose real cybersecurity risks, but a U.S. ban would not eliminate those threats and could create risks of its own by driving startups to sketchier providers.
Neill Blomkamp Releases First AI-Generated Short Film ‘Nightborne’
Director Neill Blomkamp, known for District 9, released Nightborne, his first short film generated entirely by AI. The 13-minute sci-fi horror piece was built using the Seedance 2.0 video generation model, with Blomkamp directing frame by frame through text prompts. Shot in a documentary style, the film follows a U.S. pilot presumed dead and a secret military program that sends fallen soldiers back into combat. The faces and voices of 32 real people were used under licensing agreements, and human artists created the concept art. Blomkamp said on X that he plans to shoot a full-length feature film in the same format. He founded Barley Studios, a new AI film studio that succeeds Oats Studios. AI still faces skepticism in creative industries over training data sourced without consent, job losses, copyright issues, and declining quality, but it is reportedly gaining ground as a production tool.
$400M Nonprofit Current AI Launches to Build Public Option for AI
A nonprofit called Current AI launched at France’s AI Action Summit in early 2025 with a goal of building a public option for AI, open and free like the early web. The French government seeded it with 100 million euros, joined by the Ford and MacArthur foundations, DeepMind, and Salesforce. Total commitments now top $400 million, and the group aims to mobilize $2.5 billion over five years. CEO Ayah Bdeir, who previously ran Mozilla’s AI strategy and founded littleBits, said backers are funders, not investors. Current AI’s first focus is the world’s languages, half of which are at risk of dying out while English dominates big models. In India, it worked with the government’s language unit on Suno Sutra, a pocket-sized offline device that runs AI in 22 Indian languages with no internet and is open-sourced. Last month the group handed $3.2 million in first grants to four organizations in Kenya, Lebanon, and the Brazilian Amazon for projects including datasets for 50-plus African languages, offline tools for Indigenous Amazon data, and audit tools. Current AI also launched Alpha Chat, an open-source chatbot assembled in seven weeks by ten organizations including Hugging Face, Mozilla, and MIT Media Lab. It struck a deal with Japan’s Sakana AI to build a shared open stack for Japanese and Global South communities. Bdeir acknowledged scale is not always the measure, and the bigger bet is a full open stack from chips to apps. The open question is whether a nonprofit can out-build companies giving models away for market share, but Current AI is betting that a public option, however small, is worth building anyway.
A $400M nonprofit backed by France and DeepMind is building AI’s ‘public option’ →