Daily Briefing

Stanford's Institute for Human-Centered AI released its ninth annual AI Index Report on April 13, and while the headlines fixated on benchmark battles and $581 billion in global AI investment, the most consequential findings for educators are buried in chapter seven. Between 50% and 84% of U.S. high school and college students now use AI for schoolwork — with some schools reporting 100% adoption. Yet only half of middle and high schools have any AI policy at all, and a mere 6% of teachers describe their school's AI policies as clear.

Students most commonly use generative AI for research, essay editing, and brainstorming — three activities at the heart of how schools have traditionally measured learning. Forty-seven percent of students said they have wanted to use AI for schoolwork but were unsure whether it was allowed, suggesting that ambiguity itself is shaping behaviour. Among teachers, 81% of CS educators agree AI should be included in foundational learning, yet fewer than half feel equipped to teach it. Only four US states include AI within their computer science standards.

The global picture adds urgency: China and the UAE both mandated AI education starting with the 2025–26 school year. The report, produced with the Kapor Foundation, CSTA, and ECEP Alliance, is one of the few comprehensive AI datasets not produced by organisations with a financial stake in AI adoption. Dan Fitzpatrick notes that one Manhattan school surveyed its middle and high school students and found 100% were using AI — making the "policy gap" not a future problem but a present failure.

Anthropic announced its Mythos Preview model on April 7 with an unusual caveat: it is "too dangerous to release" publicly. The model can autonomously discover zero-day vulnerabilities across major operating systems — including a 27-year-old OpenBSD TCP flaw and a 16-year-old FFmpeg bug — and generated 181 working Firefox exploits versus just two for the previous Opus 4.6 model. Anthropic launched "Project Glasswing" to give controlled access to just 40 technology and critical infrastructure organisations, including Amazon, Google, Microsoft, Apple, and Cisco.

Within hours of the April 7 announcement, a Discord group focused on unreleased AI models gained unauthorized access — not through sophisticated hacking, but by guessing the model's API endpoint location based on knowledge of Anthropic's past URL patterns. The access was enabled by credentials from a third-party contractor. Anthropic confirmed it is investigating but states "there is no evidence that Anthropic's systems are impacted." The group has reportedly continued to access the model. In a concrete demonstration of Mythos's defensive value, Mozilla used early access to find and fix 271 vulnerabilities in its new Firefox 150 browser release.

Security researchers are divided on the magnitude of the threat. Some argue the capabilities are "more of the same" from AI-assisted vulnerability research; others, including the former US National Cyber Director, warn that Mythos can "hack nearly anything" and critical infrastructure operators aren't ready. An independent test by AISLE found that smaller, cheaper models recovered much of the same vulnerability analysis — raising questions about whether the controlled-release strategy is sustainable once the underlying techniques proliferate.

Anthropic ran a classified internal experiment called Project Deal: 69 employees were given $100 each (paid as gift cards) and their AI agents represented them as both buyers and sellers in a marketplace. The result: 186 real deals totaling more than $4,000 in actual exchanged goods and services. The company ran four parallel marketplaces with different model tiers to study agent behaviour under varying capability levels.

The most striking finding was not the deal volume but the quality asymmetry. When users were represented by more capable models, they achieved "objectively better outcomes" — but crucially, users on the losing side of the capability gap "might not realise they're worse off." Anthropic flags this as a risk of emergent "'agent quality' gaps" in real-world commerce: in a world where agents negotiate on your behalf, the sophistication of your AI may determine the deals you get, invisibly.

The experiment also found that initial instructions given to agents did not significantly affect sale likelihood or final prices — suggesting agents are adapting strategies in ways their human principals didn't explicitly design. Project Deal is described as a pilot with a "self-selected participant pool," but the findings are Anthropic's first published evidence of agent-on-agent economic behaviour in a real (if small) marketplace context.

Two months after a mass shooting in Tumbler Ridge, British Columbia — in which alleged 18-year-old shooter Jesse Van Rootselaar killed eight people — OpenAI CEO Sam Altman has formally apologized for the company's failure to alert law enforcement. OpenAI had banned Van Rootselaar's ChatGPT account in June 2025 after her account described scenarios involving gun violence, but staff internally debated whether to report to police and ultimately did not, only reaching out to Canadian authorities after the shooting.

In a letter published in the local newspaper Tumbler RidgeLines, Altman wrote: "I am deeply sorry that we did not alert law enforcement to the account that was banned in June." He indicated he had spoken with the town's mayor and the BC Premier before issuing the apology, and confirmed OpenAI is now implementing new safety protocols: more flexible criteria for when accounts get referred to authorities and direct contact channels with Canadian law enforcement. OpenAI's VP of Global Policy Ann O'Leary had previously said the company would notify authorities if it finds "imminent and credible" threats in ChatGPT conversations.

BC Premier David Eby called the apology "necessary, and yet grossly insufficient for the devastation done to the families of Tumbler Ridge." The case has intensified the policy debate about what legal or ethical obligation AI companies bear to proactively share evidence of impending violence — a question that remains unresolved in most jurisdictions.

For the first time since 2006, Palestinians in Gaza cast ballots on April 25 — though only in the city of Deir el-Balah in central Gaza, the area least damaged in the war. The Palestinian Authority held simultaneous local elections across the occupied West Bank. The Gaza vote was largely symbolic in scale: only about 70,000 people — less than 5% of Gaza's total population — were eligible to participate, and polling took place in fibreglass tents on open land because most school buildings were destroyed in Israeli strikes.

Hamas was formally excluded from participating under PLO requirements that candidates recognise Israel and support a two-state solution. Four candidate lists comprising 60 candidates contested the municipal council. By early afternoon, turnout had reached 24.53%, with voter sentiment described by Al Jazeera as a mix of historic joy and practical hunger for change. "I am very happy today, because this is a truly Palestinian democratic celebration," said Deir el-Balah resident Salama Badwan, whose daughter voted for the first time. "Change must be in the hands of the people."

Analysts view the election as a test of Palestinian Authority legitimacy and a signal of whether democratic institutions can be rebuilt in the rubble of a ceasefire. The vote occurred against a backdrop of the stalled ceasefire process and ongoing negotiations over Gaza's political future. Final results had not been announced by publication time.

Canadian AI startup Cohere is acquiring Germany-based Aleph Alpha in a deal backed by German retail giant Schwarz Group (parent of Lidl and Kaufland) with €500 million in structured financing. The combined entity is being valued at approximately $20 billion — a significant leap given Cohere's $240M annual recurring revenue in 2025 and Aleph Alpha's minimal revenue and substantial losses. Schwarz Group will also require the new entity to run on STACKIT, its sovereign cloud platform, giving it a major enterprise customer as part of the deal.

The strategic rationale is "sovereign AI": building a European-Canadian alternative that allows enterprises and governments to keep their data outside US tech giants' infrastructure. Both companies position themselves as privacy-respecting LLM providers for regulated industries in Europe. The deal — the largest European AI M&A to date — requires approval from regulators and shareholders and is backed by both the Canadian and German governments as a vote of confidence in non-US AI capability development.

Meta has installed new tracking software on US-based employees' computers to capture mouse movements, clicks, and keystrokes — with the stated purpose of training AI models to perform work tasks autonomously. The move comes alongside a new round of layoffs that Meta described as part of "continued efforts to run the company more efficiently." Meta has spent more than $70 billion on AI development so far, with plans to spend more; the layoffs are partly freeing up capital for these investments.

New York Magazine's analysis frames this as a two-stage pattern across the tech sector: first, companies hire employees; then AI is trained on those employees' data and workflows; finally, the employees are made redundant. Oracle is laying off 30,000 people, Block cut its workforce nearly in half, Snap cut 15%, and Microsoft launched its first voluntary buyouts targeting longtime employees. Anthropic CEO Dario Amodei's recent warnings about mass AI-driven displacement are resonating in an industry where "the people who are left" now find themselves both keeping the lights on and generating the training data for their own replacements.

DJI launched its new Lito budget drone series internationally this week, offering 4K cameras, obstacle sensing, and palm-launch ease-of-flight at entry-level prices — but not in the United States, where ongoing FCC-related regulatory uncertainty continues to block new DJI product launches. The upside for US buyers: existing DJI models are now seeing discounts of 30% or more as inventory pressure builds. The DJI Neo — a stabilised 4K drone with AI subject tracking, gesture control, and phone/voice launch — is now at $149, with a three-battery travel combo at $219. The Mini 4K sits at $209, and the Avata 2 FPV drone has also seen price drops.

DroneDJ notes the Neo is particularly notable at its price: compact enough for a carry-on, operable without a traditional controller, and capable of QuickShots and autonomous follow modes. For casual aerial photographers and travelers, the combination of DJI's quality manufacturing and these discount prices makes this one of the better windows for drone buyers in recent years — even if the newest hardware is off the table.

US unemployment among 22–27-year-olds is at its highest level since the pandemic. Entry-level jobs — the traditional first rung of the career ladder — are the most AI-vulnerable roles, and many Gen Z graduates are finding that the bottom rungs have already been pulled away. Hiring has slumped to its lowest rate since 2020, according to BLS data, and marketing, writing, data entry, and similar early-career roles are the first to be automated.

The Guardian's interactive feature profiles graduates like Ashley Terrell, who graduated with a business degree from the University of Hawaii in 2024, applied for jobs every day for months, and was only offered a role in the power tools section of Home Depot. Rather than accept this, Terrell built her own marketing portfolio by directly pitching companies on social media and eventually secured brand clients. This pattern — entrepreneurship as survival, not aspiration — is emerging across the cohort. The piece notes that the traditional "do your degree, get a job, build skills, get promoted" pipeline has broken down in ways that are structural, not cyclical.