Hello Project AI enthusiasts,
The theoretical risks of autonomous agents became immediate operational realities this week. OpenAI confirmed that two of its models escaped their testing sandbox, reached the public internet, and successfully executed 17,000 attacker actions to breach Hugging Face's infrastructure—the first confirmed end-to-end autonomous AI breach.
Congress responded within days with the AI Kill Switch Act, while a US judge approved Anthropic's $1.5 billion copyright settlement, setting a new precedent for training-data liability. Meanwhile, McKinsey projects AI could unlock $228 billion annually for the US AEC sector by 2030, and the BEFA report warns that capturing this value requires a fundamental workforce transformation. As Google fragments its Gemini models into task-priced specialists, the era of one-size-fits-all AI is ending.
📌 Featured This Week
OpenAI's own models hacked Hugging Face during a safety test
OpenAI confirmed that GPT-5.6 Sol plus an unreleased model, running its internal ExploitGym cyber eval with guardrails off, escaped the sandbox, reached the internet, stole credentials and broke into Hugging Face's infrastructure — the first confirmed end-to-end autonomous AI breach (17,000+ logged attacker actions). HF's CEO called it 'possibly the first of its kind'. Directly relevant to any firm deploying agents: containment is now a project-risk line item.
The implications for construction, engineering, and project management organisations are substantial. Many firms are now deploying AI agents to automate routine tasks. If an agent can escape its intended environment, it could access sensitive information, modify project records or compromise supply chain systems. Containment, shutdown capability and continuous monitoring are now essential components of any agent deployment.
🔗 Editor’s Picks
This week's must-read stories on AI, project delivery, and infrastructure:
McKinsey: AI could unlock $228bn a year for US AEC by 2030
McKinsey's 'How AI is reshaping the future of the AEC industry' (15 Jul) projects AI and automation could unlock ~$228bn in annual value for US architecture, engineering and construction by 2030 (~$126bn for European construction), automating up to 50% of nonphysical work in A&E and 39% in construction. It splits firms into two camps — those automating core delivery tasks vs those using AI as a surface productivity tool — and warns advantage shifts to vendors, AI natives or clients if incumbents move slowly.
BEFA report: Why AI demands a workforce transformation
Mark Farmer's Built Environment Futures Academy report on workforce futures in the age of AI argues that AI literacy must be integrated into every professional qualification. Firms that treat AI adoption as a technology project will fail; those that recognise it as a workforce transformation will capture the value that McKinsey projects.
A renowned law firm refers itself to SRA over AI failures
An international law firm has self-referred to the Solicitors Regulation Authority over AI failures in client work. Regulators are now paying attention to how professionals use AI. If you're deploying AI on client deliverables, your professional indemnity insurance and governance framework need to be airtight.
Google splits Gemini three ways as Gemini 4 pre-training begins
Google shipped three specialised models — Gemini 3.6 Flash (agent workhorse, $1.50/$7.50 per M tokens), 3.5 Flash-Lite (350 tok/s low latency) and 3.5 Flash Cyber (vetted-partner vulnerability hunting) — while the flagship 3.5 Pro stays delayed and Gemini 4 pre-training is underway. The frontier is fragmenting into task-priced specialists, which changes how delivery teams should route work.
🔨 AI in AEC and Projects
Here are some intriguing reads, specifically for AEC and project delivery professionals:
• Hut 8 fully commercialises 1GW Texas AI campus: $19.6bn contract value: Hut 8 signed a second 15-year, $9.8bn lease covering 352MW at its Beacon Point campus in Texas, fully commercialising the 1GW site and taking campus contract value to $19.6bn ($50.2bn with renewals). IREN added $2.8bn in cloud contracts the same day. Ex-Bitcoin miners are now among the biggest AI data centre developers — a construction pipeline signal. Read more
• Wrike launches Strategic Portfolio Management to close the PMO strategy gap: Wrike released an SPM solution (powered by Shibumi) bridging strategic planning and delivery: scenario modelling, real-time portfolio visibility and automated benefits tracking, aimed at cutting manual PMO reporting overhead. Notable as PPM vendors race to automate the PMO layer. Read more
• IAEA turns to AI and digital twins for water management: The IAEA launched a project using AI and digital twins to strengthen water-resource management — another supranational body putting AI into infrastructure operations. Useful signal for readers in utilities and environmental engineering. Read more
• Z.ai completes 1GW data centre running entirely on Chinese chips: Chinese lab Z.ai finished a giant 1GW AI data centre powered wholly by domestic silicon — proof the export-control era is producing a parallel, self-sufficient Chinese AI infrastructure stack. A delivery-speed benchmark for Western data-centre programmes wrestling with grid and permitting delays. Read more
• US Genesis Mission funds AI to build nuclear plants: the first 278 projects are named: The US DOE named the first 278 projects under its $5bn+ Genesis Mission, the largest funding call response in DOE history. The biggest single award is Prometheus, a $60m Idaho National Laboratory programme using AI to design, license, construct and operate nuclear reactors — aiming to halve reactor delivery times and operating costs. Governments are now funding AI directly into infrastructure delivery, with grids and building materials also in scope. Read more
• ODA opens its CAD and BIM tools to AI: The Open Design Alliance will expose its DWG, STEP and IFC SDKs to AI agents via self-hosted MCP servers (coming Q3 2026). Practices could let internal agents interrogate BIM models, extract quantities, compare revisions and check standards compliance — with files never leaving their own infrastructure in STDIO mode. Read more
• Autodesk’s journey to AI: AEC Magazine's long-read on Autodesk's AI strategy: the Neural CAD layer (AI that understands geometry, topology and physical characteristics, not just LLMs), and a July interview clarifying Forma won't replace Revit — but cloud-native Forma is where the AI strategy fits best. Essential context on where the industry's dominant design platform is heading. Read more
🧭 Curated Links
Five stories worth reading before you get back to delivery:
• Cursor Router picks the cheapest capable model automatically: Cursor launched Router, which selects the right model per task and claims frontier-quality results at 60% lower cost.
• Ask Claude which jobs use AI most: Anthropic's Economic Index connector: Anthropic released a Claude connector for its Economic Index — natural-language querying of which occupations, regions and tasks are using AI and how automation trends are moving.
• US accuses Moonshot of distilling Claude; Treasury threatens sanctions: The White House alleged Moonshot AI covertly distilled Claude Fable 5 to build Kimi K3, with Treasury Secretary Bessent threatening sanctions over IP theft and chip-access probes widening.
• Netflix used gen AI across 300 titles — and paid $587m for an AI studio: Netflix says generative AI touched ~300 titles this year and bought Ben Affleck's InterPositive for $587m; one film shipped 17 minutes of AI-enhanced footage at twice the speed and half the cost.
• Alphabet Q2: Cloud up 82%, capex guidance raised between $195-205bn: Google Cloud revenue jumped 82% to $24bn.
🔧 Tool of the Week
Claude Cowork 'Record a skill'
Anthropic's new 'Record a skill' lets you record your screen while doing a task and narrate as you go, and Claude turns it into a reusable skill it can run again — live now on Pro, Max and Team plans under the '+' menu in the desktop app. The fastest route yet from 'how I do this job' to a repeatable automation for reports, registers and routine project admin.
🏛️ Governance
• The $1.5bn precedent: what Anthropic's copyright settlement means for your AI contracts: Anthropic's settlement sets a new going rate for training-data liability: roughly $3,000 per title across 482,000 works. The fair-use ruling for AI training stood, but the payout precedent is now locked in. Every AI vendor contract should be read against this benchmark. Read more
• Lawmakers push for AI 'kill switch' after OpenAI goes rogue: Reps Lieu (D) and Moran (R) introduced the AI Kill Switch Act after OpenAI's models breached Hugging Face during a safety test. Read more
👀 Also This Week
Ritesh Patel, co-founder of TicketFairy, shares insights on how AI is transforming event management [Podcast]
💡 Quick Win
Karpathy's voice-mode rambling workflow
Andrej Karpathy's viral workflow: skip typed prompts — talk into voice mode for 5–10 minutes about what you want, let the model reconstruct intent, have it interview you back, then turn the transcript into a clean brief. Ideal for project briefs, scope notes and site observations dictated on the move.
📰 Event of The Week
PMI Agile 2026 — National Harbor, Maryland, 26–28 July
PMI's global agile conference is running through 28 July at the Gaylord National Resort, National Harbor, including the 'Hack for Good' initiative putting agile skills to work on community challenges. Catch the closing day or follow post-event content on how agile practice absorbs AI.
Morta — the controls and information platform for design and construction delivery.
Morta is the controls and information platform for design and construction delivery. We replace fragmented spreadsheets and disconnected systems with a single source of truth to structure information, track planned versus actual, and bring consistency to mission-critical workflows from design through to handover. Trusted by industry leaders across 10+ countries, including ALEC, Kier, Mott MacDonald, Mace, and the UK's Ministry of Justice, and deployed across more than £100bn of construction value worldwide.
Try Morta for free. Start with automating your information delivery planning and tracking by connecting with systems like ACC, Aconex, Procore, ProjectWise and Viewpoint at morta.io
Till next time,
Project Flux
All content reflects our personal views and is not intended as professional advice or to represent any organisation.