UK Regains Early Model Access as Amazon Chases Decart and Sakana Tests Self-Improving Agent Teams

Global AI News Daily Ledger · Monday, 12 October 2026

Today's strongest developments concern frontier oversight, chip strategy and learning from agent teams: UK testing access is restored, Amazon reportedly pursues Decart, and Sakana introduces MASS. HCLTech expands Claude distribution while Rivet, GitHub and Hermes strengthen agent infrastructure. Commerce funding and two smaller product releases round out the edition.

1. UK regains pre-release access to frontier AI models

Reporting. The Telegraph reported on 11 October at 19:46 BST that the White House reversed its request to withhold new models from the UK's AI Security Institute. LBC corroborates the report. The change restores testing access to systems from Anthropic, OpenAI and Google DeepMind after ministerial lobbying; the earlier restriction had affected Claude Mythos 5.1.

Analysis. A material change to cross-border safety oversight, rather than a new model release. Access permits evaluation, not proof that a model is safe. This is reported government action; neither article supplies the directive itself. The original report was published after yesterday's cutoff.

Sources: Telegraph / LBC · 11-12 Oct 2026 · Original reporting · Corroborating account

2. Amazon reportedly negotiates a $7 billion Decart acquisition

Reporting. The Wall Street Journal reported on 11 October at 16:26 ET that Amazon is in talks to buy Decart for around $7 billion, citing people familiar with the matter. CTech reports that negotiations follow a failed Anthropic sale process. Decart develops software for efficient model execution across chips and real-time interactive video systems. No agreement is publicly confirmed.

Analysis. A potential link between Amazon's Trainium strategy and model-efficiency software, with world-model capability also in the mix. Treat price and negotiations as anonymous-source reporting, not an announced acquisition. Different outlets repeating the Journal are not independent confirmation of a signed deal.

Sources: Wall Street Journal / CTech · 11-12 Oct 2026 · Original deal reporting · Context and caveats

3. Sakana AI introduces MASS, a multi-agent self-improvement loop

Reporting. Sakana AI's 12 October announcement describes one Qwen3.6-27B model acting as solver, workflow optimizer and evaluator, then learning from its best team executions. After two cycles, the authors report 1.2-1.6 times the base model's score per output token on four research benchmarks. Their detailed report also shows 0.96 times on Terminal-Bench 2.0 and 0.94 times on SWE-bench Verified.

Analysis. A concrete research result on learning coordination, not evidence of unlimited recursive improvement. External model judges were used for final evaluation, not supervision inside the loop. The synthetic training suite is small, coding transfer is uneven, and shared blind spots can reinforce errors. Results are author-reported, not independently reproduced.

Sources: Sakana AI / research report / AI Weekly · 12 Oct 2026 · Dated primary announcement · Methods and negative results · Independent technical coverage

4. HCLTech and Anthropic expand governed enterprise-agent deployment

Reporting. HCLTech announced a Global Premier Partnership with Anthropic on 12 October, including a dedicated business unit and center of excellence. Its services will use Claude, Cowork and Claude Code; HCLSoftware plans selected integrations including BigFix and AppScan. Claude and Claude Code will also enter AI Force workflows for IT support and code remediation, with policy controls and human oversight for high-impact changes.

Analysis. A distribution and implementation deal connecting a frontier lab to large enterprise systems. The agreement is verified through the issuer release and CNBC TV18 reporting. Planned integrations and promises of safe deployment are not measured outcomes or evidence that the products are already rolled out to every customer.

Sources: HCLTech / CNBC TV18 · 12 Oct 2026 · Issuer announcement · Independent reporting

5. Rivet agentOS 0.4 adds remotely managed durable Hosted VMs

Reporting. Rivet's 12 October release introduces Hosted VMs created by ID and controlled through RivetKit or HTTP. Its API covers files, processes, terminals, software, networking and cron. Configuration survives restarts, output survives reconnects, and scoped JWTs permit direct access to one VM. The release adds dynamic software installation, resource limits and expiring port previews.

Analysis. Agent infrastructure aimed at reliable execution beyond one transient sandbox. Rivet's 4.8 ms median cold-start and large cost-advantage figures are vendor benchmarks for the VM itself, not end-to-end HTTP latency; network time must be added. The release and API are primary-source verified, but no independent performance validation was found.

Sources: Rivet · 12 Oct 2026 · Primary release · API contract

6. GitHub Agentic Workflows hardens sandboxes and session evidence

Reporting. GitHub's gh-aw v0.91.8 pre-release, listed at 11 October 22:33 UTC, adds unified elapsed-time session timelines across multiple agent engines, memory schema validation and better model-routing diagnostics. The notes say a custom grader sandbox escape was closed and workflow YAML, scanner handling and temporary-file writes were hardened. The 12 October weekly update confirms the release's scope.

Analysis. A developer-control update with a specific security fix, not a general guarantee of agent containment. Its pre-release status matters for adoption. The daily event is v0.91.8; older releases summarized in the same weekly post are not counted as new stories. No independent exploit analysis was located.

Sources: GitHub release / official weekly update · 11-12 Oct 2026 · Versioned release listing · Dated official summary

7. Hermes Agent 0.21.7 switches source installs to stable releases

Reporting. Nous Research's Hermes Agent v0.21.7 release listing is timestamped 12 October 05:53 UTC, although the notes call it an 11 October release. Official source installs now follow published tags rather than every main-branch commit, with a channel switch in desktop settings. Signed reasoning survives resumed chats and Bedrock Converse sessions; update checks better handle firewalls and GitHub rate limits.

Analysis. A practical reliability change for agent environments where unreviewed updates and broken resume can disrupt work. Forks and mirrors still follow main; Store, MSIX and bundled installs do not change update policy. The tag rolls up roughly 204 merged PRs, but future v0.22.0 documentation is not counted as a release today.

Sources: Nous Research / Freedom.Tech · 12 Oct 2026 · Primary release listing · Release coverage

8. Loops announces $3.5 million for an integrated commerce-agent stack

Reporting. The Next Web and SiliconANGLE report a $3.5 million seed round announced on 12 October, led by a16z speedrun with NTT Docomo and other investors. Loops combines natural-language product search, shopping assistance, catalog enrichment and virtual try-on. The company says funds will support product work, deployments and expansion rather than selling retailers disconnected AI features.

Analysis. A smaller vertical-agent funding signal, not a frontier-model event. Company-supplied customer counts and conversion gains are not independent trials; the reported 83% purchase-conversion comparison is observational. a16z's earlier investment predates today, so the fresh event is the full seed-round announcement, not a newly discovered first investor.

Sources: The Next Web / SiliconANGLE · 12 Oct 2026 · Funding and measurement caveat · Independent funding account · Investor portfolio record

9. Flozic launches a redesigned cross-app agent platform

Reporting. Flozic's 12 October announcement says its redesigned platform combines goal-driven agents, multi-step workflows, voice, chat and computer-use agents with thousands of integrations. The company says existing loops keep running without migration. Its live product page describes run history, versioning, retries and human-in-the-loop checkpoints alongside autonomous execution.

Analysis. A vendor launch in the crowded workflow-agent market, retained as a smaller builder signal. Yahoo's article is PRNewswire distribution, not independent reporting. Integration counts, deployment speed, compliance labels and reliable autonomy are vendor claims, not findings from an audit. Cross-app action still needs scoped permissions and checks on each consequential effect.

Sources: Flozic / issuer release via PRNewswire · 12 Oct 2026 · Dated issuer release · Current product claims

10. MorningAI adds a website-building agent with review before publication

Reporting. MorningAI's 12 October product update adds a Website Agent that reads a connected GitHub site, plans edits and creates pages, UI changes and images from plain language. The vendor says changes appear in a desktop, tablet and mobile preview and go live only after the user publishes. The same update adds social campaign planning and scheduling, plus email and web analytics in Channels.

Analysis. A smaller vibe-coding and marketing-agent release with an explicit review boundary. The publication gate is more meaningful than a claim of hands-free site management, but it still requires checking the actual changes. This is primary-source product news; no independent test of output quality, security or publishing controls was found.

Sources: MorningAI · 12 Oct 2026 · Dated product update · Agent product documentation

Edition note. Exactly ten distinct events are included. Smaller vendor launches are labeled and not presented as major frontier breakthroughs. The coverage window begins at the previous edition's 18:48 IST cutoff; date-only 12 October announcements are within the new day. No exact public X post URL was recovered, so there is no Firsthand signal entry. Older Anthropic incident disclosures, Mellum2.1, Liquid d1, AWS Physical AI Toolchain and Poppy were excluded despite new coverage dates. OpenCode 2.0.27 could not be reconciled with the official changelog and was omitted. Vendor benchmarks and anonymous-source deal reporting remain attributed. Prior editions are unchanged.