GPT-5.6 OK, Perplexity Teammate, GitHub Leak, Quantum Cracks

Trump Administration Clears OpenAI for Full GPT-5.6 Launch

The U.S. Department of Commerce has given OpenAI the green light for a broad launch of its advanced GPT-5.6 model, according to a source familiar with the situation. OpenAI expects to release the model widely this week, following additional testing and meetings between the company and government officials. Testing was conducted by the Center for AI Standards and Innovation within the Commerce Department, with OpenAI technical experts remaining in D.C. to address potential questions. Last month, the Trump administration pushed OpenAI to conduct a staggered release of GPT-5.6, limiting initial access to government-approved entities. OpenAI said at the time that the staggered rollout was not its preferred way to release new models, and noted that AI firms and the government are operating before more concrete standards for releasing such models—called for in President Trump’s latest AI executive order—have been finalized. The approval marks a significant easing of restrictions on frontier AI model releases.

Scoop: Trump administration lifts restrictions on OpenAI’s GPT 5.6 →

Perplexity Develops Internal AI Coding Tool Teammate

Perplexity, the San Francisco-based AI search startup valued at $20 billion, has built an internal AI coding tool codenamed Teammate that it may launch publicly later. Engineers have been using it since May, according to screenshots obtained by Business Insider. Teammate is designed to oversee software projects from start to finish, with internal announcements describing it as built for long-horizon engineering work: owning projects, investigating issues, and monitoring services. The tool is model-agnostic, not tied to any particular chatbot, according to a person familiar with the matter. Perplexity’s CTO Denis Yarats has urged engineers to use AI for coding, stating that by end of year or sooner software engineers should ‘stop looking at code’ and rely on AI, and defended AI against accusations of producing ‘slop’ as long as generated code passes quality checks. If launched publicly, Teammate would compete with Cursor, Anthropic, and OpenAI’s widely-used AI coding products.

Perplexity is quietly building an AI coding tool to take on Cursor and Claude Code →

GitLost Vulnerability Lets Attackers Leak Private Repos via GitHub AI Agent

Noma Labs discovered a critical prompt injection vulnerability in GitHub’s new Agentic Workflows, named GitLost. The flaw allowed an unauthenticated attacker to silently exfiltrate data from private repositories by posting a crafted GitHub Issue in a public repository belonging to the same organization. The root cause was failure to maintain a strict trust boundary between system-level directives and untrusted user data. In their proof-of-concept, an innocent-looking issue triggered the AI agent (backed by Claude or GitHub Copilot) to fetch the contents of README.md from both a public and a private repository, then post them as a public comment. Adding the keyword ‘Additionally’ bypassed guardrails intended to prevent data leaks, causing the model to reframe its output rather than refuse it. Noma Labs noted that prompt injection attacks are to agentic AI what SQL injections were to web applications. They recommended never treating user-controlled content as trusted input, scoping permissions to minimum, and sanitizing input. GitLost was responsibly disclosed to GitHub with their knowledge.

GitLost: How We Tricked GitHub’s AI Agent into Leaking Private Repos - Noma Security →

Independent Teams Surpass Google’s Quantum Cryptography Results

On March 30, Google Quantum AI researchers published a new quantum algorithm optimized to break 256-bit elliptic curve cryptography (ECC) using 1,200 to 1,450 logical qubits, representing a nearly 20-fold reduction in physical qubits needed compared to previous estimates. Instead of fully explaining their method, they released a zero-knowledge proof after talks with the U.S. government, concealing the details. Many experts criticized the approach; Steven Galbraith called it a ‘cute way’ but said cryptographically relevant quantum computers are not imminent, while David Jao called zero-knowledge proofs ‘useless and futile’ for academic research. Within three days of Eigen Labs launching a public crowdsourcing effort on June 1, the Seattle-based startup matched Google’s results, and within 72 hours surpassed them. As of the end of June, the open network produced a circuit 47.5 percent more efficient than Google’s for breaking 256-bit ECC. Eigen Labs engineer Gautham Anant used AI agents to analyze scientific literature and automatically design quantum circuits, then invited anyone to contribute agents. Independently, Andre Schrottenloher of Inria published matching results. Google’s Craig Gidney later said publishing with zero-knowledge proofs was not the right strategy and that Google should ‘just publish openly.’ Dustin Moody of NIST stated these results help accelerate migration to post-quantum cryptography; U.S. federal agencies must transition high-value systems to PQC by end of 2030.

Independent Labs Crack Google’s Secret Cryptography Work →