AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
This was the week frontier AI became more useful, more expensive and harder to govern.
The standout release was Anthropic’s Claude Fable 5, the first publicly available Mythos-class model. The claim is not just better chat. Anthropic is positioning it around long-horizon software engineering, complex analytical work, vision, scientific research and million-token context. The caveat matters: sensitive cyber, biology, chemistry and model-distillation requests can be routed away from Fable 5 to Opus 4.8, and business traffic on Mythos-class models is subject to 30-day retention for safety monitoring. For most business users, that trade-off is acceptable, but it should not be ignored.
Apple’s WWDC reset was also significant, but more strategically than practically. Siri AI moves Apple closer to the assistant model everyone expected years ago: personal context, on-screen awareness and cross-app actions. The direction is right, but availability, device limits and regional delays mean this is not yet something most UK businesses can operationalise this week.
The broader pattern is clearer. AI capability is still moving quickly, but so are the constraints around it: compute cost, safety gating, data-retention trade-offs, model access tiers and procurement risk. The sensible response is not to chase every release. It is to test the few tools that could materially change workflow quality, while putting basic controls around confidentiality and spend.
For UK leaders, the practical takeaway is simple: try Fable 5 before the included-access window closes, and turn on stronger ChatGPT protections for anyone using AI with confidential work.
What to Try This Week
Try Claude Fable 5 before 22 June.
This is the one release worth making time for. Anthropic says Fable 5 is included at no extra cost on Pro, Max, Team and seat-based Enterprise plans until 22 June, after which it moves to usage credits unless capacity allows an extension.
Do not test it on generic prompts. Give it one difficult piece of real work that normally takes concentration and context: a messy diligence document, a product spec, a complex spreadsheet explanation, a codebase issue, a board-paper rewrite, or a market-map synthesis. Ask it to show assumptions, identify gaps, and produce a usable first draft.
The point is not whether it sounds clever. The test is whether it gets you materially closer to a decision or deliverable than your current model.
Turn on ChatGPT Lockdown Mode for confidential work.
This is less exciting than a new model, but probably more important for day-to-day business use. If your team uses ChatGPT for client documents, legal correspondence, market research, board papers or anything commercially sensitive, Lockdown Mode is a sensible default.
It reduces exposure to indirect prompt-injection attacks by restricting network behaviour, including internet image fetching, file downloads, Deep Research and Agent Mode. That will break some workflows, so use it where confidentiality matters most rather than across everything blindly.
A good rule of thumb: if the input would make you uncomfortable sitting in someone else’s inbox, use the more restricted mode.
Geopolitics, Governance and Big Moves
SpaceX prices the year’s biggest IPO at around $135 a share
SpaceX is expected to price its IPO at around $135 a share this week ahead of trading on Friday 12 June, against a $1.75 trillion target valuation. The Information’s pre-IPO subscriber Q&A on Wednesday set the pricing-week schedule. Pricing-week conditions are not gentle: the Nasdaq has fallen 5% over the past week, with interest-rate and Iran-war concerns layered on top. Goldman Sachs, the lead bank, told investors it expects SpaceX revenue to reach $474 billion by 2030 from $18.7 billion last year, two-thirds of it from AI. Martin Peers flagged the projection as harder to credit than OpenAI’s own $284 billion 2030 forecast: the xAI line item Goldman is using to justify the AI two-thirds had only $3.2 billion of 2024 revenue, substantially all of it from X advertising, which has dropped by half since the Musk acquisition.
OpenAI files confidential S-1, declares its “third phase”
OpenAI filed a confidential S-1 with the SEC on Monday 8 June. The company explicitly hedged on timing: “it may be a while because there are things we want to do that are likely easier as a private company.” Sam Altman and chief scientist Jakub Pachocki published a same-day blog titled “Built to benefit everyone,” declaring OpenAI has entered its third phase after pure research and product shipping: “the economy is now being shaped around AI.” Three stated goals: an automated AI researcher, accelerating the economy while sharing gains widely, and giving “everyone on Earth a personal AGI.” A noteworthy line, “entirely automating everything is not the future we want,” echoes Anthropic’s recent pause framing. OpenAI also proposed a global coordination body and launched an Economic Research Exchange for external researchers studying AI’s economic effects.
Anthropic now generates revenue 35% above OpenAI
The Information’s same-day “Anthropic vs. OpenAI” brief made the new revenue gap explicit: Anthropic is now generating revenue at an annualised rate roughly 35% higher than OpenAI, a sharp reversal from late 2025 when Anthropic was generating less than half. Anthropic has discussed going public as soon as Q4 and projects free-cash-flow positive by 2028. OpenAI projects to burn more than $100 billion before potentially reaching free-cash-flow positive around 2030. Its $20 billion-plus chip agreement with Cerebras already signals the scale of that burn. Perplexity separately said it plans to IPO in 2028 “regardless of what happens to Anthropic or OpenAI,” the first time a second-tier AI lab has put a specific public-listing date on its own roadmap.
Apple confirms Siri AI at WWDC, running on Google Gemini
Apple opened WWDC 2026 on Monday 8 June with the long-trailed Siri reset, rebranded Siri AI and now running on Apple’s own models in collaboration with Google Gemini. The assistant gets its own dedicated app, chatbot-style interactions with access to past conversations, screen-context awareness, file uploads, and the ability to take systemwide app actions. It will arrive as a free fall update for iPhone 15 Pro and newer with a public beta next month, but it will not launch in the EU or China. The most powerful on-device model and its features will require iPhone 17 Pro or iPhone Air, and several AI features ship with daily-usage limits that push users to iCloud subscriptions. Martin Peers’ read: “the bar was low.” Federighi used the keynote to take a public shot at competitors “racing forward, seemingly pursuing AI for the sake of AI.” Apple also announced iOS 27, a new on-device Core AI framework giving developers Swift APIs across CPU, GPU and Neural Engine, and confirmed John Ternus as the incoming CEO from 1 September.
Bots have overtaken humans on the internet, a year early
Cloudflare CEO Matthew Prince confirmed on Thursday 4 June that bot traffic now accounts for between 52 and 62% of daily internet traffic, averaging 57.4% bots to 42.5% humans over the past week. Prince’s previous public forecast was late 2027, so the milestone has arrived roughly a year early, driven by fast-growing agentic AI traffic rather than search-engine crawling. The country-level distribution is striking: Gibraltar leads at 92.1% bot traffic, followed by Singapore (76.3%), Iran (76.2%), Ireland (72.8%) and the Netherlands (68.8%). For UK B2C operators, agent-first user journeys (who arrives, what they extract, whether they can act) are now the strategic question, not the edge case. Display advertising, attribution and anti-abuse systems were all built around a human-majority baseline that no longer holds.
Four-lab bioweapons letter is the first cross-frontier policy alignment of 2026
The CEOs of Anthropic, OpenAI, Google DeepMind and Microsoft co-signed an open letter pressing US Congress to require synthetic-DNA sellers to vet every buyer and order, warning that frontier AI now enables non-experts to design bioweapons. This is the first time Sam Altman and Dario Amodei have publicly co-signed the same policy document in 2026 and the first time all four frontier-lab CEOs have aligned on a regulatory ask. UK Department for Science, Innovation and Technology now has to choose whether to back the request. The trans-Atlantic pre-deployment partnership architecture sketched in May with CAISI is the most natural vehicle.
Tech, Tools and Releases
Anthropic releases Claude Fable 5, bringing Mythos-class capability to the public
Anthropic launched Claude Fable 5, its first publicly available Mythos-class model, alongside the more powerful Claude Mythos 5 for trusted partners. Fable 5 brings frontier capability into broader use while restricting higher-risk cybersecurity and bioweapons use cases. Early testers describe it as extremely strong, but slower and expensive, with standout coding, reasoning and long-context performance, including a 1 million-token context window and 128,000-token outputs.
Apple uses WWDC to reset Siri for the AI era
Apple used WWDC 2026 to unveil Siri AI, a major rebuild of its assistant around conversational interaction, personal context and cross-app task execution. The new system draws on Apple Intelligence, updated foundation models and collaboration with Google Gemini, while retaining Apple’s privacy-heavy on-device positioning. Siri can understand emails, photos and on-screen activity, and act across apps. Apple also announced iOS 27, stronger parental controls and Image Playground upgrades.
OpenAI ships Lockdown Mode, the first consumer-tier defence against indirect prompt injection
OpenAI rolled out Lockdown Mode this week, an optional ChatGPT setting that gives stricter protection against indirect prompt-injection attacks, where hidden malicious instructions on web pages or in documents trick the model into leaking data. Turning it on disables Deep Research and Agent Mode, blocks ChatGPT from pulling or displaying internet images, and stops the chatbot from downloading files, though manual uploads and on-platform image generation still work. The mechanism is network-layer rather than parser-layer: Lockdown Mode limits the network requests an attacker could exploit, rather than trying to detect injection patterns in processed content. The feature is available to all personal accounts including the free tier. This is the first mainstream consumer-tier defence designed specifically for the indirect prompt-injection class of attack and lands inside the same fortnight as Anthropic’s open-sourced Defending Code Reference Harness.
Anthropic publishes “When AI builds itself”: 80% of new production code now written by Claude
Anthropic published “When AI builds itself” on Thursday 4 June, the first frontier-lab long-form report grounded in internal recursive-self-improvement metrics rather than hypothetical scenarios. The headline numbers: Claude now authors 80% of Anthropic’s new production code and engineers ship 8x more code per year compared with 12 months ago. Anthropic uses the report to argue for a global slowdown or temporary pause on AI development, comparing the coordination requirement to nuclear-weapons treaties. The report names MiniMax, whose new M2.7 model “helped build itself”, as evidence that recursive self-improvement is now a multi-lab pattern across the US and China. Critics quoted in TLDR AI framed the pause call as a marketing move to slow rivals. Either reading lands in the same place: recursive self-improvement is the new centrepiece of the safety debate, and Anthropic’s framing is the document that subsequent moves will reference.
Cognition rebrands Windsurf as Devin Desktop, pitches the “Switzerland of AI agents”
The Information’s Rocket Drew reported on Sunday 7 June that Cognition is making over its app to position itself as a neutral hub for AI coding agents, rebranding Windsurf as Devin Desktop. The framing, explicitly “the Switzerland of AI agents”, places Cognition as the agent-orchestration layer that does not pick favourites between OpenAI Codex, Anthropic Claude Code or Google Antigravity, in deliberate contrast to OpenAI’s announcement that it would merge Codex and ChatGPT into a single agentic surface. Cognition followed up on Tuesday with FrontierCode, the first benchmark to measure code mergeability, whether a model’s output can actually be merged into a live codebase without breaking things. For UK CTOs evaluating coding-agent platforms with teams using agents from three different vendors, the Cognition pitch is the cleanest current alternative to vendor lock-in.
Xiaomi MiMo-UltraSpeed lands a 1,000 tokens-per-second claim on commodity hardware
TLDR AI led its Tuesday “Headlines & Launches” with Xiaomi’s release of MiMo-V2.5-Pro-UltraSpeed, a 1-trillion-parameter language model developed with inference partner TileRT and clocking 1,000 tokens per second on a standard 8-GPU commodity node, roughly 15x the median frontier-model rate cited for ChatGPT and Claude. The speedup is achieved through FP4 quantisation on the model’s expert layers and DFlash speculative decoding. The model is available through a limited API trial from 9 to 23 June at three times the standard rate for roughly ten times the output. For UK enterprise teams, a 1-trillion-parameter Chinese model with credible 1,000 tokens-per-second throughput on commodity hardware is the cleanest case yet for treating Chinese frontier models as a serious cost-control lever, even before the gray-market access stories that have dominated W23 and W24 reporting.
Instagram AI chatbot breach hits 20,225 accounts
Meta confirmed this week that hackers compromised at least 20,225 Instagram accounts by exploiting a flaw in its AI-powered “High Touch Support” account-recovery chatbot, the first time Meta has put a number on the scope of the breach. The bug: the chatbot sent password reset links to any email address without checking whether it actually belonged to the Instagram account. Attackers exploited it from 17 April to 31 May 2026. Potentially exposed data: contact info, birth dates, posts, direct messages and profile information. Meta disabled the chatbot, invalidated reset links and placed affected users into a mandatory security checkpoint. This is the first quantified consumer-AI safety failure of 2026 with a confirmed account count and a counterweight to the consumer-tier defence story of OpenAI Lockdown Mode. AI is shipping consumer security improvements and consumer security failures in the same week.
Quick Hits
The “SaaSpocalypse” is over, but public stocks haven’t noticed. Private equity mogul Orlando Bravo declared the call on CNBC on Tuesday and Marc Benioff retweeted it. Salesforce is down 34% year-to-date, ServiceNow down 30% and Asana down 45%. The SpaceX, Anthropic and OpenAI IPO window has not yet absorbed capital out of weaker tech stocks into the higher-growth listings.
AI Boomerang: 32% of laid-off roles are being rehired. A Robert Half survey of 2,000 US hiring managers shows 32% have eliminated a role because of productivity gains from AI or automation, only to later rehire for the exact same position. Finance leads at 44%, then HR at 35%, then tech at 32%. Managers cited “gaps in quality, oversight and decision-making once business demands picked up after the initial cuts.”
Anthropic embeds engineers at the NSA, names Glasswing coalition. The Financial Times reported Anthropic has embedded around six engineers at the NSA to deploy its Mythos model for offensive cyber operations targeting networks in countries such as China or Iran. The report explicitly names Microsoft, Apple and Amazon as Project Glasswing coalition members for the first time.
Anthropic cuts off xAI’s access to its products. The Information’s TITV show confirmed production-level enforcement of the W24 reporting that xAI trained its coding models on Claude outputs for months before being cut off. This is the cleanest single example of a frontier lab enforcing its terms of service against a US peer, landing inside the SpaceX IPO pricing window.
China announces a $295 billion AI buildout plan. Bloomberg reports a 2-trillion-yuan five-year blueprint to link computing sites into one grid by 2028, with Huawei and other local suppliers providing at least 80% of the core technology. The UK announced a £1.1 billion AI hardware plan the same day.
Apple Poke approved as the first AI agent on Messages for Business. Apple opened the iMessage agent surface to a third party for the first time, days before WWDC.
We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.