Section Technology
X agrees to geo-block UK visibility of proscribed terrorist accounts under Ofcom undertakings
15 May 2026 filings under the Online Safety Act include a 48-hour triage pledge for suspected illegal terrorist and hate material plus escalation paths for specialist reviewers when automated queues misfire.

Ofcom published 15 May 2026 undertakings from X under the Online Safety Act, including a commitment to block United Kingdom IP ranges from viewing profiles tied to Home Office–proscribed terrorist organisations. The company also pledged to triage reports of suspected illegal terrorist and hate posts within 48 hours and to route edge cases to subject-matter reviewers—an operational detail that matters when classifiers mis-label news or open-source conflict documentation as glorification.
Compliance engineering must reconcile geo-fences with VPN paths, quote-post chains that re-broadcast banned URLs, and third-party API embeds on publisher sites. Ofcom can levy large fines for systematic breaches, so risk registers now sit beside CFO-level capital planning—not only trust-and-safety spreadsheets.
What the commitment table covers
| Layer | Regulator expectation |
|---|---|
| Account visibility | Hard UK visibility block for proscribed entity profiles |
| Discovery | Remove or down-rank shortcuts that trivially surface banned propaganda |
| User reports | 48-hour triage clock from submission; appeals cannot dead-end in bot-only loops |
UK terrorism statutes already criminalise narrow forms of glorification; platforms argue ranking is not endorsement, while prosecutors test how far algorithmic surfacing can imply culpable states of mind.
Free-expression and journalism friction
Archives of Arab Spring footage, Syrian civil-defence media, and academic OSINT threads sometimes collide with hash databases. Ofcom guidance encourages context panels, but false positives that hide war-crimes evidence remain a live reputational and legal risk for distributors.
Election-year political speech fights sit on a different track from illegal-terror duties—yet both share the same content stack, so product teams must segment policies to avoid over-broad takedowns.
Commercial and cross-border spillover
Brand-safety audits already pressure advertiser lists; EU Digital Services Act transparency templates may converge with UK reporting, nudging multinationals toward single engineering builds with jurisdiction flags.
SSL transparency and subdomain provisioning sometimes reveal sandbox environments where classifiers are tested before production rollout—signals analysts watch, but they remain indirect evidence of policy change.
What would reset the compliance read next
Quarterly Ofcom transparency statistics, High Court reviews of geo-block challenges, appeal outcome ratios for terrorism flags, and advertiser contract clauses tying spend to safety metrics would each supply hard numbers beyond the undertaking text.
National reporting portals remain the correct channel for violent extremist material in the UK—re-sharing primary propaganda feeds verification work to the wrong side of the ledger.
Suggested reading
Other stories that pair well with this one—often from the same section or on overlapping themes.
Google CLI Links OpenClaw to Gmail Unsupported
Google's open-source Workspace CLI on GitHub links AI agents including OpenClaw to Gmail and Drive, but the company labels the project unsupported and warns workflows may break as APIs evolve.
Google I/O 2026 Pushes Always-On Gemini Agent
Google I/O 2026 in Mountain View spotlighted Gemini Spark, described as an always-on personal agent across Workspace and other apps—with user approval before sensitive actions—plus faster Gemini models, agentic Search, and Android XR hardware.
Claude Code Auto Mode routes risky tool calls through a Sonnet 4.6 classifier instead of endless taps
Anthropic’s March 2026 engineering deep dive frames Auto Mode as permission automation: a two-stage transcript filter plus a prompt-injection probe, built after internal telemetry showed users accepting 93% of manual prompts anyway.
Anthropic’s Q1 2026 growth reads near 80× in markets coverage; Semi Analysis tallies put ARR above $44 billion
Benzinga and syndicated Fortune copy captured chief executive Dario Amodei calling the pace “too hard to handle” around an 80-fold quarterly surge narrative, while a Semi Analysis digest summarized by trade press puts annualized run-rate revenue above $44 billion after a climb from about $9 billion at year-end 2025.
Revolut rolls out a physical Dogecoin-branded card in the U.K. and wider EEA
The neobank’s first crypto-culture plastic works on Visa and Mastercard rails, pairs with Apple Pay and Google Pay in supporting setups, and leans on fiat balances even as the artwork leans on DOGE memes; Own The Doge licensing framed charity tie-ins in launch copy.
Anthropic buys Stainless, the API-to-SDK toolchain rivals including OpenAI and Google relied on
The 2022 New York startup led by former Stripe engineer Alex Rattray automated libraries across Python, TypeScript, Kotlin, Go, and Java; Anthropic confirms it will wind down hosted products for other vendors while letting past customers keep generated code.
Walmart’s six new Onn Android 16 tablets from $97: spec sheet, who they beat, and who should skip them
Launch-day listings describe Android 16 across the stack—from a 7-inch Helio G80 starter through a 13-inch Pro bundle with stylus—but paper wins still need reality checks against Amazon’s Fire line, Lenovo’s budget slabs, and discounted Samsung Tab hardware.
UK AI Security Institute publishes Mythos Preview cyber scores: 73% on expert CTFs, first model to finish a 32-step range in three of ten runs
AISI’s 13 April 2026 write-up summarises controlled evaluations of Anthropic’s Claude Mythos Preview on capture-the-flag tasks and on “The Last Ones,” a 32-step simulated corporate intrusion; Opus 4.6 remains the nearest comparator on the multi-step range but trails on step count.
Oakland jury shuts Musk’s OpenAI fight on a clock question, not the ‘betrayed lab’ plot
Nine Northern District jurors agreed the February 2024 filing landed outside the limitations window they were instructed to use; Judge Yvonne Gonzalez Rogers still formalises the advisory result, but the merits of charitable-trust and enrichment theories never went to a second-phase verdict.
Calif’s Mythos-on-M5 kernel exploit story gains an official Apple footnote in macOS Tahoe 26.5 security credits
Calif still narrates seven-day lab work with Memory Integrity Enforcement on macOS 26; Apple’s catalogue page for Tahoe 26.5 now lists CVE-2026-28952 as reported by Calif.io in collaboration with Claude and Anthropic Research—a narrower confirmation than Calif’s full chain narrative but stronger than silence.
Keep exploring
Browse the full archive or return to the front page.
Sources and external links
Sources and filings our editors consulted to verify this story. External links open in a new tab.