A single HTTP request could hijack any unpatched Ruflo AI agent
Noma Security's CVSS-10.0 disclosure traces the hole to Ruflo's own Docker defaults, and the fix stops new attacks without undoing damage already done.
Ruflo's MCP bridge let one HTTP request run shell commands1
Ruflo bills itself, and Noma Security's disclosure repeats, as the leading open-source AI agent meta-harness, citing roughly 10 million downloads and 1 million active users. Those figures come from the firm that found and is publicizing the flaw, not from an independent audit.2 The flaw itself is not in dispute: Ruflo's own GitHub Security Advisory confirms its MCP Bridge shipped with no authentication on the default Docker Compose configuration's POST /mcp endpoint.1 The bridge exposed 233 tools, per Noma's count, including ruflo__terminal_execute, so a single unauthenticated HTTP request could run arbitrary shell commands inside the container.2 The advisory puts the flaw at CVSS 10.0, affected versions below 3.16.3 and patched in 3.16.3, and confirms the terminal-execute blocklist enforced in the autopilot flow never covered direct calls to the bridge.1 Its remediation guidance, rotating API keys and auditing the AgentDB memory store, makes clear the patch stops new exploitation without undoing compromise that happened before it.1 The Hacker News's independent account of the same advisory carries that distinction too, plus an on-the-record confirmation from maintainer Reuven Cohen.3
Gemini Robotics 2 unscrews a lightbulb 92% of the time, screws one in just 36%4
DeepMind's own numbers for Gemini Robotics 2, the three-model suite it introduced 30 July, split sharply by task. Unscrewing a lightbulb succeeds 92% of the time, screwing one back in just 36%, using a dustpan 32%, and sealing a ziplock bag 40%. Precise insertion reaches 89.6% and picking items off a shelf 76.3%.4 The suite separates whole-body manipulation (the vision-language-action model, Gemini Robotics 2) from multi-step planning and multi-robot coordination (Gemini Robotics ER 2).4 A third, on-device model adapts to a new robot body from fewer than 200 demonstrations with no network access.4 The company's own post states plainly that "multi-finger dexterous manipulation remains challenging."4 ER 2 is available through Google AI Studio and, in private preview, the Gemini Enterprise Agent Platform. The VLA and on-device models remain limited to early-access partners and trusted testers.4 TheNextWeb, reviewing the same release independently, traces the split to hardware: Apollo 2's five-fingered, 22 degree-of-freedom hand draws the low multi-finger scores, while simpler two-fingered grippers on other platforms cluster at 74% to 90%. It adds that the robots "pause to think through moves a person makes without a second thought."5
The EU opens bidding for seven AI "gigafactories" worth over €30 billion6
The European Commission opened its formal call for AI Gigafactory proposals on 30 July. The offer: up to €10 billion in EU and national public funding, at least €20 billion in private investment, over €30 billion combined, across as many as seven sites.6 Bids are due 12 November 2026, with award decisions expected in early 2027.6 Euronews's independent breakdown of that €10 billion finds only about €1 billion currently committed. Roughly €5 billion is anticipated from a future EU budget member states have not yet agreed, and another €5 billion is expected from the states themselves, a distinction the Commission's own release does not carry.7 The program's stated goal is letting Europe "develop advanced AI on its own infrastructure, in line with EU rules and values." Euronews reports memoranda already signed with Nvidia, AMD, and Qualcomm, none of them EU chip designers.7
A cryogenic controller ran quantum error correction with no room-temperature help8
HRL Laboratories built a cryo-CMOS controller that operates at 4 Kelvin and draws no more than 3.5 watts. It ran error correction directly on the lab's 18-qubit silicon processor, using a distance-5 repetition code and a quantum error-detecting code, with no real-time help from room-temperature electronics.8 That result comes from the team's preprint posted to arXiv, the same text later published in Nature behind a paywall the preprint does not share.8 The chip holds 54 exchange-coupled quantum dots configurable into 18 exchange-only qubits. The paper reports a mean single-qubit gate error of 2×10⁻⁴ and a mean two-qubit (CNOT) gate error of 3×10⁻³, with the best reproducible CNOT run reaching 9×10⁻⁴.8 The authors describe that performance as advancing the exchange-only state of the art "by an order of magnitude."8 Nature's own news desk, independent of HRL, reports the result published the same day as a comparable one from QuTech in Delft, on a separate 5-qubit device. Silicon spin qubits still trail superconducting systems (100-plus qubits) and neutral-atom systems (thousands) in raw scale, even after this gain.9
Sources
- GitHub Security Advisory · Unauthenticated RCE in ruflo MCP bridge default docker-compose deployment (GHSA-c4hm-4h84-2cf3)
- Noma Security · RufRoot: The MCP Bridge Vulnerability That Turns Agents Into Rogue Admins (CVE-2026-59726)
- The Hacker News · Ruflo MCP Flaw Lets Unauthenticated Attackers Run Commands and Poison AI Memory
- Google DeepMind · Gemini Robotics 2 brings whole body intelligence to robots
- TheNextWeb · Google's new AI can now walk a humanoid across the room and tidy up, feet to fingertips
- European Commission · EU launches AI Gigafactories call to boost Europe's computing capacity and unlock more than €30 billion in investment
- Euronews · EU opens call for seven "gigafactories" to train next-generation AI technologies
- arXiv · A digitally controlled silicon quantum processing unit
- Nature · Underdog "spin qubits" leap forward in race to a useful quantum computer