OpenAI ships Dreaming V3 memory and Lockdown Mode; Microsoft debuts 7 in-house models at Build 2026; Great American AI Act enters legislative process.

Run: 02–07 Jun 2026 (5-day window; strict 72h yielded 5 items, expanded per signal-density rule) · 28 items reviewed → 10 published · 4 verified · 6 secondary · 0 rumor · 42% exploration · Run timestamp: 2026-06-07

TL;DR

• OpenAI Codex for knowledge workers — Sites and role-specific plugins expand Codex from developers to six enterprise job categories, integrating 62 business apps out of the box. (→ OpenAI Codex knowledge work)
• ChatGPT Dreaming V3 — background memory synthesis replaces manual saved-memories lists; rolling out to Free tier. (→ ChatGPT Dreaming V3)
• OpenAI Lockdown Mode — new opt-in security setting disables web access, Agent Mode, and live connectors to defend against prompt injection. (→ OpenAI Lockdown Mode)
• Microsoft MAI models — 7 proprietary models including MAI-Thinking-1 (reasoning) and MAI-Code-1-Flash (coding) launched at Build 2026 without OpenAI training dependency. (→ Microsoft MAI model family)
• Great American AI Act — bipartisan discussion draft proposes 3-year federal preemption of state AI laws and mandatory audits for large frontier developers. (→ Great American AI Act)

Items

OpenAI Codex adds Sites, six role-specific plugins, and 62 business-app integrations

Source: https://openai.com/index/codex-for-knowledge-work/ · OpenAI · 2026-06-02 Verification: T2 secondary · announcement · workflow-automation / dev-tools Tier note: Primary URL returned HTTP 403. TechCrunch (2026-06-02), Axios (2026-06-02), and VentureBeat (2026-06-02) independently confirm the date and feature details.

OpenAI extended Codex beyond software engineering on 2 June 2026, introducing Sites (Codex-generated interactive websites hosted by OpenAI), six role-specific plugins targeting data analytics, creative production, sales, product design, equity investing, and investment banking, and an Annotations feature for precise in-document commands. The six plugins aggregate 62 popular business applications — including Snowflake, Figma, and Salesforce — with 110 automated skills pre-wired. Knowledge workers now represent about 20% of Codex’s 5 million weekly active users and are growing more than three times as fast as the developer segment.

Why it matters for automation/productivity: Teams that have avoided Codex as a developer-only tool now have pre-built connectors to the business apps they already use, lowering the integration barrier for non-technical roles to deploy agent-based workflows. The Sites feature delivers Codex work product as a shareable hosted web app rather than a local file, enabling lightweight distribution of agent-generated analysis or dashboards without engineering overhead.

Key claims:

• 5 million weekly active users, ~20% non-developer → secondary (TechCrunch, Axios)
• 62 business apps, 110 automated skills → secondary (VentureBeat)
• 6 role categories: data analytics, creative, sales, product design, equity investing, investment banking → secondary (multiple outlets)

Cross-references:

• https://techcrunch.com/2026/06/02/openai-launches-new-codex-tools-for-white-collar-work/ (T2, corroborating)
• https://www.axios.com/2026/06/02/openai-codex-knowledge-workers (T3, corroborating)
• https://venturebeat.com/orchestration/openais-codex-update-lets-agents-build-interactive-enterprise-workspaces-via-sites-and-role-specific-plugins (T3, corroborating)

Caveats: Plugin rollout is phased; not all roles may have full access at the announced date. Sites are hosted by OpenAI — organizations with strict data residency requirements should evaluate before deploying work product to hosted Sites.

OpenAI frontier models and Codex reach AWS Bedrock general availability

Source: https://openai.com/index/openai-frontier-models-and-codex-are-now-available-on-aws/ · OpenAI · 2026-06-02 Verification: T2 secondary · announcement · dev-tools / workflow-automation Tier note: OpenAI primary URL returned HTTP 403. Amazon confirmed via www.aboutamazon.com/news/aws/bedrock-openai-models (T2, AWS official news page).

GPT-5.5, GPT-5.4, and Codex became generally available on Amazon Bedrock on 2 June 2026, covering commercial and GovCloud regions. Enterprises access the models through AWS’s existing IAM, procurement, billing, and compliance frameworks without requiring separate OpenAI accounts or API agreements. The launch also introduced Amazon Bedrock Managed Agents powered by OpenAI, enabling agentic workflows inside AWS-native infrastructure.

Why it matters for automation/productivity: Organizations already operating inside the AWS ecosystem can now deploy OpenAI models without routing traffic or billing through OpenAI’s own platform, removing procurement, data-residency, and compliance barriers that have blocked enterprise adoption. Agents built on Bedrock inherit AWS CloudTrail audit logs, IAM access control, and data governance — the same stack organizations use for production AWS workloads.

Key claims:

• GPT-5.5, GPT-5.4, and Codex on Bedrock GA → AWS official (aboutamazon.com, T2)
• Available in commercial and GovCloud regions → AWS confirmation
• Bedrock Managed Agents powered by OpenAI → secondary (multiple T3 sources)

Cross-references:

• https://www.aboutamazon.com/news/aws/bedrock-openai-models (T2, corroborating — AWS official)
• https://techcrunch.com/2026/06/02/openai-launches-new-codex-tools-for-white-collar-work/ (T3, contextual)

ChatGPT Dreaming V3 replaces saved-memory lists with background synthesis

Source: https://openai.com/index/chatgpt-memory-dreaming/ · OpenAI · 2026-06-04 Verification: T2 secondary · announcement · productivity-ai Tier note: Primary URL returned HTTP 403. 9to5Mac (2026-06-04), TechTimes (2026-06-05), and Windows News (2026-06-05) independently confirm date and feature details.

OpenAI deployed Dreaming V3 on 4 June 2026, replacing ChatGPT’s manually curated saved-memories list with a background process that synthesizes user context from past conversations without prompting. Memories auto-update over time — a future-dated travel note becomes a past-dated recollection after the event — and are visible in a reviewable summary page where users can add, edit, or delete entries. The rollout began with US Plus and Pro subscribers; Free-tier expansion follows, enabled by a roughly 5x reduction in compute cost for running the synthesis process at scale.

Why it matters for automation/productivity: Recurring work contexts — ongoing project details, communication style preferences, client-specific constraints — can now persist across ChatGPT sessions without manual setup, reducing onboarding friction for repetitive professional tasks. Teams using ChatGPT for recurring client workflows benefit most, provided users periodically audit the memory summary for stale or incorrect entries.

Key claims:

• 5x compute cost reduction enabling Free tier rollout → OpenAI-claimed; secondaries confirm (9to5Mac, TechTimes)
• Plus and Pro users first; Free-tier expansion follows → 9to5Mac, TechTimes
• Memory summary page, user-editable → multiple secondary sources
• Factual recall 82.8%, preference adherence 71.3%, time-sensitive accuracy 75.1% → OpenAI internal evaluation, not independently replicated

Cross-references:

• https://9to5mac.com/2026/06/04/openai-says-chatgpts-memory-feature-is-getting-smarter-and-coming-to-free-users/ (T3, corroborating)
• https://www.techtimes.com/articles/317840/20260605/chatgpt-memory-dreaming-update-openai-rewrites-personalization-engine-limits-audit-trail.htm (T3, corroborating — surfaces GDPR profiling concern)

Caveats: Performance metrics (82.8% factual recall) are OpenAI internal figures, not independently replicated. EU GDPR classifies persistent behavioral synthesis as profiling, triggering consent obligations and erasure rights. OpenAI has not disclosed how Dreaming V3 interacts with EU AI Act transparency obligations effective August 2026.

OpenAI Lockdown Mode disables ChatGPT tool access to defend against prompt injection

Source: https://openai.com/index/introducing-lockdown-mode-and-elevated-risk-labels-in-chatgpt/ · OpenAI · 2026-06-04 Verification: T2 secondary · announcement · dev-tools Tier note: Primary URL returned HTTP 403. TechCrunch (2026-06-06) and The Hacker News (2026-06-04) independently confirm the date and feature scope.

OpenAI began rolling out Lockdown Mode on 4 June 2026 as an opt-in security setting available across all personal ChatGPT accounts (Free, Go, Plus, Pro) and self-serve Business accounts. When enabled, it disables live web access, image rendering in responses, Deep Research, Agent Mode, Canvas networking, live connectors, and file downloads — the surface area through which prompt injection can arrive via external content. A companion Elevated Risk label appears when Lockdown Mode is off and a connected tool is processing content from an external source.

Why it matters for automation/productivity: Enterprises deploying ChatGPT in workflows that process documents or web content from untrusted sources now have a built-in option to reduce prompt injection exposure without custom guardrails. The Elevated Risk label provides in-session security awareness for users who cannot tolerate full lockdown — a practical middle path for most business contexts that balance functionality against risk.

Key claims:

• Available on Free, Go, Plus, Pro, and self-serve Business accounts → TechCrunch (T2), The Hacker News (T3)
• Disables: web access, image responses, Deep Research, Agent Mode, Canvas, connectors, file downloads → multiple secondaries
• Elevated Risk label for connected tools processing external content → TechCrunch

Cross-references:

• https://techcrunch.com/2026/06/06/openai-unveils-lockdown-mode-to-protect-sensitive-data-from-prompt-injection-attacks/ (T2, corroborating)
• https://thehackernews.com/2026/06/new-chatgpt-lockdown-mode-limits-tools.html (T3, corroborating)

Caveats: OpenAI states Lockdown Mode does not guarantee complete protection — prompt injections can still arrive via cached web content or uploaded files. This is a UX-layer risk reduction measure, not a cryptographic security guarantee.

Claude Code v2.1.166 severs cross-session authority to harden multi-agent security

Source: https://github.com/anthropics/claude-code/releases · Anthropic · 2026-06-06 Verification: T2 verified · changelog · dev-tools

Claude Code shipped v2.1.166 on 6 June 2026 with a targeted security change: messages relayed via SendMessage from other Claude sessions no longer carry user authority. Previously, a session receiving instructions from a separate Claude session could act with full user permissions, creating a potential lateral privilege escalation path in multi-agent orchestration setups. The release also added fallback model configuration, expanded glob pattern support in deny rules, improved thinking controls, and fixed JetBrains IDE terminal flickering. An earlier same-week release (v2.1.162, 4 June) added waitingFor status output showing what a waiting session is blocked on, fixed WebFetch domain permission handling, and quieted startup messaging.

Why it matters for automation/productivity: Teams running multi-agent Claude Code workflows where one orchestrator session instructs subordinate sessions now have an explicit authority boundary — a subordinate cannot elevate its permissions via inter-session messaging. This removes a category of privilege escalation risk relevant to production agentic deployments, particularly those exposing Claude Code to external inputs.

Key claims:

• Cross-session SendMessage messages no longer carry user authority → GitHub releases primary (T2)
• v2.1.162: waitingFor status for blocked sessions → GitHub releases primary (T2)

Microsoft launches 7-model MAI family at Build 2026, trained independently from OpenAI

Source: https://blogs.microsoft.com/blog/2026/06/02/microsoft-build-2026-be-yourself-at-work/ · Microsoft Blog · 2026-06-02 Verification: T2 verified · announcement · model-release / dev-tools Tier note: Comparative benchmark claims (MAI-Thinking-1 vs. Claude Sonnet) downgraded to T4 — Microsoft-run evaluations, no independent replication at launch. Simon Willison’s independent review (T2) found the ‘clean commercially licensed data’ framing overstated; training included a proprietary web crawl of ~1.2 trillion pages and Common Crawl data.

Microsoft announced seven in-house AI models under the MAI brand at Build 2026 (2–3 June, San Francisco): MAI-Thinking-1 (35B active parameters, 1T total, 256K context, reasoning), MAI-Code-1-Flash (137B parameters, 5B active, coding), MAI-Image-2.5 (text-to-image and image-to-image), MAI-Transcribe-1.5 (43 languages), and MAI-Voice-2 (15+ languages). MAI-Code-1-Flash began rolling out to GitHub Copilot Free, Pro, Pro+, and Max plans on 2 June at $0.75/M input and $4.50/M output tokens. MAI-Thinking-1 entered private preview on Azure Foundry.

Why it matters for automation/productivity: Organizations using GitHub Copilot now have a Microsoft-trained coding model available alongside OpenAI models, providing a diversification option. Enterprises concerned about single-vendor dependency in their AI stack can treat the MAI family as a signal that Microsoft is building an internal supply chain, even though independent validation of capability claims is not yet available.

Key claims:

• 7 models in the MAI family → Microsoft Blog (T2)
• MAI-Code-1-Flash: $0.75/M input, $4.50/M output → GitHub Changelog 2026-06-02 (T2)
• MAI-Thinking-1: 35B active / 1T total parameters, 256K context → Microsoft Blog (T2)
• Training used proprietary web crawl (~1.2T pages) and Common Crawl → Simon Willison (T2), contradicts Microsoft’s ‘clean licensed data’ framing

Cross-references:

• https://github.blog/changelog/2026-06-02-mai-code-1-flash-is-now-available-for-github-copilot/ (T2, corroborating — confirms availability date)
• https://simonwillison.net/2026/Jun/2/microsofts-new-models/ (T2, contradicting — data sourcing and parameter count corrections)

Caveats: MAI-Thinking-1’s preference claim vs. Sonnet 4.6 is from Microsoft’s own blind evaluation — not independently replicated. Simon Willison later corrected his initial analysis after consulting the technical paper, finding the ‘commercially licensed data’ framing applied only to a subset of the training corpus.

Anthropic expands Project Glasswing to 150 organizations covering power, water, and healthcare

Source: https://www.anthropic.com/news/expanding-project-glasswing · Anthropic · 2026-06-02 Verification: T2 verified · announcement · ai-for-business COI: Anthropic is the sole source on program scale and results. The 10,000-flaws figure has not been independently audited.

Anthropic extended Project Glasswing on 2 June 2026 to approximately 150 new organizations across 15+ countries, adding power, water, healthcare, communications, and hardware sectors not represented in the initial cohort. Partners receive gated access to Claude Mythos Preview — Anthropic’s highest-capability model, held in controlled research preview because of its offensive cyber capabilities. The initial cohort reportedly uncovered more than 10,000 high- or critical-severity flaws; Anthropic states a successful attack on most new partners would affect more than 100 million people.

Why it matters for automation/productivity: The expansion signals the operating model Anthropic is using for frontier AI with offensive capabilities: gated access tied to a specific defensive use case, not general availability. Organizations adjacent to critical infrastructure — cloud providers, government contractors, regulated utilities — should monitor Glasswing as a reference for what governed access to capability-limited AI looks like in practice, particularly as similar access-control questions arise in enterprise AI deployments.

Key claims:

• 150 new organizations in 15+ countries → Anthropic primary (T2)
• Sectors: power, water, healthcare, communications, hardware → Anthropic primary
• 10,000+ high/critical severity flaws from initial cohort → Anthropic-claimed, COI noted, not independently verified
• Successful attack on new partners would affect 100M+ people → Anthropic’s assessment

Cross-references:

• https://techcrunch.com/2026/06/02/anthropic-scales-claude-mythos-to-critical-infrastructure-in-15-countries/ (T2, corroborating)
• https://cyberscoop.com/anthropic-project-glasswing-expansion-critical-infrastructure-claude-mythos/ (T2, corroborating)

Caveats: Claude Mythos Preview is not available for general deployment. Glasswing access is invitation-only; no public application process announced. The 10,000-flaws figure is Anthropic’s own assessment without independent methodology disclosure.

White House EO on AI promotes innovation and creates voluntary frontier model access framework

Source: https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/ · The White House · 2026-06-02 Verification: T1 verified · government policy · policy-regulation

The White House signed an executive order on 2 June 2026 directing federal agencies to reduce regulatory friction on AI development while strengthening cybersecurity of AI-adjacent systems. The EO sets 30-day deadlines for CISA to issue AI-enabled cyber defense guidance and Treasury to establish an AI cybersecurity clearinghouse, and a 60-day mandate for NSA, Treasury, and CISA to develop a classified benchmarking process identifying high-capability models the EO terms ‘covered frontier models.’ A voluntary framework invites AI developers to provide government early model access before release; the EO explicitly bars mandatory licensing, preclearance, or permitting requirements for model development.

Why it matters for automation/productivity: The covered-frontier-model classification process could affect how quickly new AI capabilities clear review for use in regulated sectors — if a model is designated, commercial release timelines may shift. The explicit bar on mandatory preclearance reduces a regulatory risk that had been factored into some enterprise AI procurement roadmaps.

Key claims:

• 30-day CISA guidance deadline → Whitehouse.gov (T1)
• 60-day classified benchmarking process mandate → Whitehouse.gov (T1)
• Voluntary early access framework, no mandatory licensing → Whitehouse.gov (T1)

Great American AI Act draft proposes 3-year federal preemption of state AI laws

Source: https://obernolte.house.gov/media/press-releases/obernolte-trahan-release-discussion-draft-great-american-ai-act · Rep. Obernolte · 2026-06-04 Verification: T2 secondary · legislative discussion draft · policy-regulation Tier note: Primary URL returned HTTP 403. Confirmed via FedScoop (T2, 2026-06-04) and Roll Call (T2, 2026-06-04).

Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA) released a bipartisan discussion draft on 4 June 2026 proposing the first comprehensive federal AI governance framework for the US. Key provisions: a 3-year sunset preemption of state laws specifically regulating AI model development (not deployment or use), mandatory independent audits every 6 months for companies with more than $500M in AI revenue, $300M over 3 years to formally establish the Center for AI Standards and Innovation (CAISI) in statute, and a Labor Department AI Workforce Research Hub to study workforce displacement. The preemption explicitly does not cover state laws governing the use or deployment of AI systems.

Why it matters for automation/productivity: If enacted, the bill would consolidate the patchwork of state AI regulations into a single federal compliance target for the development phase, reducing multi-state legal overhead for organizations that build and fine-tune models. The mandatory audit requirement for large frontier developers could increase vendor compliance costs and may eventually affect API pricing. The discussion draft is early-stage; formal introduction and committee markup would follow the stakeholder feedback period.

Key claims:

• 3-year preemption of state AI development laws → FedScoop (T2), Roll Call (T2)

• $500M revenue threshold for large frontier developers → FedScoop

• $300M for CAISI over 3 years → FedScoop
• 6-month mandatory audits for large frontier developers → FedScoop

Cross-references:

• https://rollcall.com/2026/06/04/bipartisan-ai-draft-proposes-three-year-preemption-of-state-laws/ (T2, corroborating)
• https://fedscoop.com/bipartisan-great-american-ai-act-draft-proposes-new-federal-ai-governance-framework/ (T2, corroborating)
• https://www.techtimes.com/articles/317903/20260606/federal-ai-regulation-bill-freezes-state-consumer-protections-three-years-sparks-revolt.htm (T3, contradicting — consumer advocacy pushback on preemption scope)

Caveats: Discussion draft only — not yet formally introduced. Consumer advocacy groups and state attorneys general have raised concerns about the preemption scope freezing existing state consumer protections. Outcome uncertain; watch for stakeholder feedback period before formal bill introduction.

OpenAI expands ChatGPT advertising to UK Free and Go users

Source: https://releasebot.io/updates/openai · OpenAI (via Releasebot) · 2026-06-04 Verification: T3 secondary · product update · ai-for-business Tier note: No dedicated OpenAI primary page found for the UK expansion. Multiple T3 aggregators confirm the June 4 date. Original US launch (May 5) is T2 confirmed via openai.com/index/our-approach-to-advertising-and-expanding-access/.

OpenAI extended ChatGPT advertising to Free and Go users in the United Kingdom on 4 June 2026, following the 5 May US launch. The UK rollout maintains the same model as the US: ads appear at the bottom of ChatGPT answers for unpaid tiers only; Pro, Business, and Enterprise subscriptions remain ad-free. OpenAI’s claim that ads do not influence answer content has not been independently verified at US or UK launch.

Why it matters for automation/productivity: Organizations with UK employees using ChatGPT on personal Free or Go accounts should factor sponsored results into their guidance on AI-assisted workflows. For enterprises evaluating ChatGPT Business for employee-facing deployments, the UK rollout reinforces the ad-free guarantee on paid plans.

Key claims:

• UK Free and Go users receiving ads as of June 4 → Releasebot (T3), multiple aggregators
• Paid plans (Pro, Business, Enterprise) remain ad-free → consistent with May 5 US announcement

Cross-references:

• https://openai.com/index/our-approach-to-advertising-and-expanding-access/ (T2, original US announcement 2026-05-05)

Caveats: No independent verification that ads do not influence answer content. UK regulatory context (ICO, CMA) may generate scrutiny given concurrent GDPR-adjacent questions raised by Dreaming V3 rollout.

Dropped

Items considered but not published, with reason:

Limitations

• Window expansion applied: Strict 72h window (Jun 4–7) yielded 5 verified items, below the standard density target of 10–15. Window expanded to 5 days (Jun 2–7) per skill protocol. Items dated before Jun 2 are excluded even if they appeared in search results.
• OpenAI primary URLs blocked: openai.com/index/codex-for-knowledge-work/, openai.com/index/chatgpt-memory-dreaming/, openai.com/index/introducing-lockdown-mode-and-elevated-risk-labels-in-chatgpt/, and openai.com/index/openai-frontier-models-and-codex-are-now-available-on-aws/ all returned HTTP 403. All four items were confirmed via ≥2 independent T2-T3 secondary sources. Tier downgraded from T2 verified to T2 secondary.
• Congressional primary URL blocked: obernolte.house.gov/media/press-releases/ returned HTTP 403. Great American AI Act item confirmed via FedScoop (T2) and Roll Call (T2).
• Login-walled coverage: X timelines, Instagram, LinkedIn private feeds, and Discord were not accessed. Public X posts indexed by search engines were captured via site:x.com searches. Items requiring logged-in X access may have been missed.
• Microsoft MAI benchmark claims unverified: MAI-Thinking-1’s Sonnet-preference claim is from Microsoft’s own evaluations. No independent third-party benchmark comparing MAI-Thinking-1 to current-generation frontier models was available at run time. Simon Willison’s review (T2) flagged data sourcing overstated in Microsoft’s marketing framing.
• OpenAI Dreaming V3 metrics unverified: The 82.8% factual recall, 71.3% preference adherence, and 75.1% time-sensitive accuracy figures are from OpenAI’s internal evaluation. No independent replication was available.
• Anthropic Glasswing 10,000-flaws figure unverified: This figure comes from Anthropic’s own program reporting. No third-party security audit methodology or confirmation was available.
• SEA-region and cross-language coverage: An explicit Indonesian-language and broader SEA-region search found no in-window product launches from local AI vendors. The AI Ready ASEAN expansion was workforce training, not tooling — excluded as low BD relevance. The geographic gap is structural (registry is US/EU heavy).
• No research-papers items: Hugging Face Daily Papers and arXiv searches for Jun 2–7 found Holo3.1 (local computer use, Jun 2) and KVarN (KV-cache quantizer, Jun 2), but both had limited BD actionability given the absence of production deployment paths. Dropped.
• No mcp-ecosystem items: PulseMCP search found no new named MCP server launches in the Jun 2–7 window with verifiable primary sources. The broader MCP ecosystem (now 16,000+ indexed servers) continues growing, but no individual server launch rose to the publication bar this period.

Search log (compact)

Total searches: 35, of which 15 exploratory or adversarial (43%).

Suggested next runs

• OpenAI Dreaming V3 EU compliance audit — The Aug 2026 EU AI Act transparency deadline arrives while Dreaming V3 is mid-rollout; watch for ICO or CNIL guidance on whether persistent behavioral synthesis requires opt-in consent notices.
• Great American AI Act stakeholder response — Consumer advocates and state AGs are organizing against the preemption clause; track the feedback period for signals on whether the bill will be narrowed before formal introduction.
• MAI-Thinking-1 independent benchmark — Microsoft’s internal evaluation of MAI-Thinking-1 has no third-party replication; watch LMSYS Arena and public eval harnesses for independent data once the private preview opens more broadly.
• OpenAI Codex Sites data governance — Sites are OpenAI-hosted; enterprise legal and compliance teams will need to determine whether Codex-generated work product served through Sites is subject to OpenAI’s data retention terms.