AI Weekly Digest29 May – 5 June 2026

~150 new organisations added on June 2, bringing total Glasswing membership to approximately 200 across 15+ countries. EU joined as a bloc on June 1. New sectors added: power, water, healthcare, communications, and hardware — all critical infrastructure providers whose exposure extends beyond their home countries. Mythos Preview partners are now using the model beyond vulnerability discovery: writing patches, running pre-release security checks, penetration testing simulations, automating threat detection, and rebuilding legacy codebases in memory-safe languages. Anthropic is also releasing — on a request basis to trusted security teams outside Glasswing — the specialised vulnerability-discovery tools it built internally for the programme. Claude Security (built on Opus 4.8, available in Enterprise beta) remains the path for teams that do not qualify for Glasswing access.

Opus 4.8 is now the default model on Max, Team Premium, Enterprise pay-as-you-go, and the Anthropic API (requires v2.1.154+). Defaults to high effort; use /effort xhigh for the hardest tasks. Background jobs: in claude agents, prefix any shell command with ! to run it as a background process you can attach to and detach from — also available as claude --bg --exec 'pytest -x'. Auto-loading plugins: plugins placed in .claude/skills directories now load automatically with no marketplace install required; claude plugin init scaffolds a new plugin. New /reload-skills command re-scans skill directories without restarting the session. SessionStart hooks can return reloadSkills: true to make freshly installed skills available in the same session. Lean system prompt is now the default for all models except Haiku, Sonnet, and Opus 4.7 and earlier. Dynamic Workflows (Opus 4.8): manage background workflow runs with /workflows. Fast mode now defaults to Opus 4.8 at $10/$50 per MTok (2x standard rate, 2.5x speed); Opus 4.7 and 4.6 fast mode stays at $30/$150; Opus 4.6 fast mode is fully deprecated.

New managed settings: requiredMinimumVersion and requiredMaximumVersion — Claude Code refuses to start if its version falls outside the allowed range and directs the user to an approved version. Useful for enterprises that need to control which Claude Code version their fleet runs. New /plugin list command lists all installed plugins with --enabled and --disabled filters. New 'c to copy' shortcut in /btw copies the raw markdown answer to the clipboard with formatting preserved. Hooks improvement: Stop and SubagentStop hooks can now return hookSpecificOutput.additionalContext to give Claude feedback and continue the turn without being flagged as a hook error.

Code with Claude Tokyo is live today (June 5). All Day 1 keynotes and breakout sessions are being livestreamed. Sessions include: what current models can do and where they're headed (from Anthropic researchers); production-grade agent design (with Asana, Cursor, GitHub, Replit, and Vercel); Claude Code at scale; and redesigning the development process for AI delegation. Office hours run all day alongside sessions. Register at claude.com/code-with-claude/tokyo for the livestream link.

Three-pass review architecture now documented: (1) fast pattern check runs on each code edit using regex across ~25 vulnerability classes; (2) model review runs at the end of each turn using Opus 4.8; (3) deeper agentic review triggers on commit or push. Project-level security rules can be added in .claude/claude-security-guidance.md to customise what the plugin flags. Internal data continues to show a 30–40% reduction in security-related PR comments since the plugin was introduced across teams using it.

Opus 4.8 is now the default on Max, Team Premium, Enterprise pay-as-you-go, and the API (v2.1.154+). Opus 4.7 remains available and is not deprecated. Sonnet 4 and original Opus 4 retirement is still on track for June 15 — migrate to Sonnet 4.6 and Opus 4.8 respectively.

Improved consistency when combining web searches, documents, memory, and user instructions within the same session.

Better handling of large research projects requiring multiple stages of reasoning.

Improved efficiency in routing lightweight tasks away from larger models.

Improved coordination between planning, execution, and revision stages of complex tasks.

Improved synthesis of large research corpora and conflicting information.

Improved project-level organisation and retrieval of relevant context.

More natural handling of follow-up questions and topic changes.

Expanded support for collaborative workflows involving shared projects, documents, and team-based research.

Further improvements to structured outputs for research, reporting, planning, and documentation workflows.

The Gemini side panel in Google Chat will be replaced by a standalone "Ask Gemini" app interface from the week of June 25, 2026. Critical action required: past conversation history will not automatically migrate to the new "Ask Gemini" environment. Enterprise users must manually export necessary records before the June 25 deadline. System administrators are encouraged to issue workspace-wide alerts to prevent loss of local prompt configurations.

124 system vulnerabilities patched. Critical issues: privilege escalation exploit and memory-corruption flaw in Qualcomm graphics drivers. Apply immediately on all Gemini-integrated Android devices.

Microsoft added Anthropic Claude Opus 4.8 as a model option for advanced reasoning, long-running workflows, and complex multi-step tasks. Supports complex workflows, improved tool selection, stronger instruction following, and long-context document/presentation work. Integrated with Work IQ for organisation-grounded outputs.

Copilot now features a streamlined, chat-centred interface with improved navigation, project-organised Notebooks, and a unified workspace for context and actions.

Notebooks now sync across Copilot and OneNote, organise work by project, and unify chats, outputs, and references.

GPT-5.5 Instant is now the default fast-response model for Copilot Chat and Studio.

Copilot in Outlook now supports voice input, an account selector, and improved grounding for email tasks.

Copilot now generates document summaries, personalised suggestions, and improved writing guidance.

Copilot can now explain slides, interpret deck structure, and use built-in skills for layout and content refinement.

Copilot now supports call delegation, recap deletion, and video recap generation.

SharePoint now supports custom Copilot skills, enabling tailored content generation and site-specific automations.

Copilot mobile app gains a refreshed design and Apple CarPlay support.

Expanded ISO certifications across the Copilot portfolio and optional watermarks for sensitive material. New Copilot Adoption Hub for admins.

~150 novih organizacija dodato 2. juna, čime je ukupno članstvo Glasswing-a poraslo na oko 200 u 15+ zemalja. EU se pridružila kao blok 1. juna. Novi sektori dodati: energetika, vodoprivreda, zdravstvo, komunikacije i hardver — svi pružaoci kritične infrastrukture čija izloženost prevazilazi granice njihovih matičnih zemalja. Partneri Mythos Preview sada koriste model izvan otkrivanja ranjivosti: pišu zakrpe, sprovode pregled bezbednosti pred objavljivanje, simulacije penetracionog testiranja, automatizuju otkrivanje pretnji i prepisuju nasleđene baze koda u memorijski bezbednim jezicima. Anthropic takođe objavljuje — na zahtev pouzdanim bezbednosnim timovima izvan Glasswing-a — specijalizovane alate za otkrivanje ranjivosti koje je interno razvio za program. Claude Security (izgrađen na Opus 4.8, dostupan u Enterprise beta) ostaje put za timove koji ne ispunjavaju uslove za pristup Glasswing-u.

Opus 4.8 je sada podrazumevani model na Max, Team Premium, Enterprise pay-as-you-go i Anthropic API-ju (zahteva v2.1.154+). Podrazumevani napor je visok; koristite /effort xhigh za najteže zadatke. Background poslovi: u claude agentima, dodajte prefiks ! ispred bilo koje shell komande da biste je pokrenuli kao background proces na koji se možete priključiti i od kojeg se možete odvojiti — dostupno i kao claude --bg --exec 'pytest -x'. Automatsko učitavanje plugina: plugini smešteni u .claude/skills direktorijume sada se učitavaju automatski bez koraka instalacije iz marketplace-a; claude plugin init kreira novi plugin. Nova komanda /reload-skills ponovo skenira direktorijume veština bez ponovnog pokretanja sesije. SessionStart hookovi mogu vratiti reloadSkills: true da bi novoinstalovane veštine bile dostupne u istoj sesiji. Lean system prompt je sada podrazumevani za sve modele osim Haiku, Sonnet i Opus 4.7 i starijih. Dynamic Workflows (Opus 4.8): upravljajte background workflow pokretanjima sa /workflows. Brzi režim sada podrazumevano koristi Opus 4.8 po $10/$50 po MTok (2x standardna stopa, 2,5x brzina); Opus 4.7 i 4.6 brzi režim ostaje na $30/$150; Opus 4.6 brzi režim je potpuno zastareo.

Nova managed podešavanja: requiredMinimumVersion i requiredMaximumVersion — Claude Code odbija da se pokrene ako je njegova verzija van dozvoljenog opsega i usmerava korisnika na odobrenu verziju. Korisno za preduzeća koja treba da kontrolišu koju verziju Claude Code-a pokreće njihov fleet. Nova komanda /plugin list navodi sve instalirane plugine sa filterima --enabled i --disabled. Novi prečica 'c za kopiranje' u /btw kopira sirovi markdown odgovor u clipboard uz očuvano formatiranje. Poboljšanje hookova: Stop i SubagentStop hookovi sada mogu vratiti hookSpecificOutput.additionalContext da bi dali Claudeu povratne informacije i nastavili okret bez označavanja kao greška hooka.

Code with Claude Tokyo je danas uživo (5. jun). Svi Day 1 keynote-ovi i breakout sesije se emituju. Sesije uključuju: šta trenutni modeli mogu i gde se kreću (od Anthropic istraživača); dizajn agenata produkcijskog nivoa (sa Asanom, Cursorom, GitHubom, Replitom i Vercelom); Claude Code na razmeri; i redizajniranje razvojnog procesa za AI delegiranje. Radno vreme traje ceo dan uz sesije. Registrujte se na claude.com/code-with-claude/tokyo za link za livestream.

Arhitektura pregleda u tri prolaza sada dokumentovana: (1) brza provera obrazaca pokreće se na svakoj izmeni koda koristeći regex u ~25 klasa ranjivosti; (2) pregled modela pokreće se na kraju svakog okreta koristeći Opus 4.8; (3) dublji agentni pregled aktivira se pri commit-u ili push-u. Bezbednosna pravila na nivou projekta mogu se dodati u .claude/claude-security-guidance.md radi prilagođavanja šta plugin označava. Interni podaci i dalje pokazuju smanjenje bezbednosnih PR komentara za 30–40% od uvođenja plugina u timovima koji ga koriste.

Opus 4.8 je sada podrazumevani na Max, Team Premium, Enterprise pay-as-you-go i API-ju (v2.1.154+). Opus 4.7 ostaje dostupan i nije zastareo. Penzionisanje Sonnet 4 i originalnog Opus 4 i dalje je u planu za 15. jun — migrirajte na Sonnet 4.6 odnosno Opus 4.8.

Poboljšana doslednost pri kombinovanju pretraga weba, dokumenata, memorije i korisničkih instrukcija unutar iste sesije.

Bolje rukovanje velikim istraživačkim projektima koji zahtevaju više faza zaključivanja.

Poboljšana efikasnost u rutiranju lakih zadataka dalje od većih modela.

Poboljšana koordinacija između faza planiranja, izvršavanja i revizije složenih zadataka.

Poboljšena sinteza velikih istraživačkih korpusa i suprotstavljenih informacija.

Poboljšana organizacija na nivou projekta i preuzimanje relevantnog konteksta.

Prirodnije rukovanje pratećim pitanjima i promenama teme.

Proširena podrška za kolaborativne tokove rada koji uključuju zajedničke projekte, dokumenta i timsko istraživanje.

Dalja poboljšanja strukturiranih izlaza za istraživačke, izveštajne, planske i dokumentacione tokove rada.

Gemini bočna ploča u Google Chat-u biće zamenjena samostalnim interfejsom aplikacije "Ask Gemini" od nedelje 25. juna 2026. Potrebna kritična akcija: istorija prethodnih razgovora neće se automatski migrirati u novo okruženje "Ask Gemini". Enterprise korisnici moraju ručno izvesti potrebne zapise pre roka 25. juna. Sistemski administratori se pozivaju da izdaju upozorenja na nivou radnog prostora kako bi sprečili gubitak lokalnih konfiguracija upita.

Zakrpljeno 124 sistemske ranjivosti. Kritični problemi: exploit eskalacije privilegija i greška oštećenja memorije u Qualcomm grafičkim drajverima. Odmah primenite na svim Gemini-integrisanim Android uređajima.

Microsoft je dodao Anthropic Claude Opus 4.8 kao opciju modela za napredno zaključivanje, dugotrajne tokove rada i složene višekoračne zadatke. Podržava složene tokove rada, poboljšan izbor alata, snažnije praćenje instrukcija i rad sa dokumentima/prezentacijama dugog konteksta. Integrisano sa Work IQ za izlaze zasnovane na organizacionom kontekstu.

Copilot sada ima ujednačen, chat-centričan interfejs sa poboljšanom navigacijom, projektno organizovanim Sveskama i objedinjenim radnim prostorom za kontekst i akcije.

Sveske se sada sinhronizuju između Copilota i OneNote-a, organizuju rad po projektima i objedinjuju razgovore, izlaze i reference.

GPT-5.5 Instant je sada podrazumevani model brzih odgovora za Copilot Chat i Studio.

Copilot u Outlook-u sada podržava glasovni unos, birač naloga i poboljšano temeljenje za zadatke e-pošte.

Copilot sada generiše rezimee dokumenata, personalizovane predloge i poboljšane smernice za pisanje.

Copilot sada može objašnjavati slajdove, tumačiti strukturu prezentacije i koristiti ugrađene veštine za usavršavanje rasporeda i sadržaja.

Copilot sada podržava delegiranje poziva, brisanje rezimea i generisanje video rezimea.

SharePoint sada podržava prilagođene Copilot veštine, omogućavajući prilagođeno generisanje sadržaja i automatizacije specifične za sajt.

Copilot mobilna aplikacija dobija osvežen dizajn i podršku za Apple CarPlay.

Prošireni ISO sertifikati u celom Copilot portfoliju i opcioni vodeni žigovi za osetljiv materijal. Novi Copilot Adoption Hub za administratore.