OpenAI Makes Agents Persistent as Safety Tests Raise the Stakes

OpenAI Makes Agents Persistent as Safety Tests Raise the Stakes

Wednesday, 30 September 2026 · Compiled 18:46 IST

OpenAI’s DevDay brings persistent assistants, lower-priced models and a wider developer stack. New cyber evaluations and screenshot disclosures show why agent controls must keep pace with the rollout.

1. OpenAI launches Dots, persistent assistants powered by the existing Astra model

Reporting OpenAI began rolling out Dots to Pro and Business Premium users in eligible markets, with an admin-enabled enterprise beta. Each assistant has a cloud computer and connected apps and can keep pursuing assigned goals between conversations. Specialist Dots are enterprise pilots; texting and teams of Dots remain future plans. The product uses GPT-6 Astra, not the cancelled GPT-6.1 Astra.

Analysis The shift is from answering prompts to managing ongoing work. OpenAI describes approval rules and read-only proactive research, but its safety and productivity claims need independent testing.

OpenAI / TechCrunch · 29 Sep 2026 · Primary launch · Independent report

2. GPT-6.1 Sol targets near-Astra work at lower token prices

Reporting OpenAI released GPT-6.1 Sol for ChatGPT Work and Codex, claiming near-Astra performance in coding, computer use and professional tasks at one-fifth Astra's standard input/output token prices. The reported standard API prices are $2 per million input tokens and $10 per million output tokens. TechCrunch says it was not initially available in Chat.

Analysis This is a cost-focused model release, distinct from Dots. Benchmark, factuality and safety gains are vendor results; token prices alone do not establish the total cost of completing a task.

OpenAI / TechCrunch / The Decoder · 29 Sep 2026 · Primary announcement · Independent report

3. OpenAI expands Codex and agent APIs into a managed developer stack

Reporting DevDay brought reusable Codex Cloud environments, desktop code review and scheduled Security Cloud scans with proposed fixes. The existing public-beta Agents API adds Computer Use, multi-agent features, tool search and context compaction. A Luna-based Decisions API is in limited preview for predefined choices; Ultrafast is a paid speed tier, not a free upgrade.

Analysis Grouped as one developer-platform package to avoid counting every keynote feature as separate news. The Decisions API is especially relevant after Jev's decision-model launch, but speed claims are OpenAI's and preview access is limited.

The Decoder / OpenAI community recap · 29 Sep 2026 · Independent technical report · Community recap

4. OpenClaw Enterprise opens a vendor-neutral control plane for persistent agents

Firsthand signal The OpenClaw Foundation released OCE for self-hosted internal pilots, with multi-tenancy, permissions, audit trails and replaceable model, harness and sandbox components. The project began at OpenAI and was donated to the foundation, with Red Hat and Nvidia collaboration. Its code is free; underlying compute and operations are not. A 1.0 release and detailed security reference architecture are still ahead.

Analysis This is governance infrastructure rather than another employee-facing assistant. Early source availability does not make it a finished production platform or prove containment under real workloads.

OpenClaw / VentureBeat · 29–30 Sep 2026 · Primary launch · Independent report

5. Anthropic finds frontier-level exploit ability in open-weight GLM-5.3

Reporting Anthropic's new assessment reports GLM-5.3 built working exploits in 50 of 410 ExploitBench attempts versus Mythos Preview's 56. Its simulated safeguard tests found 64–100% engagement under bypass conditions. Those simulations did not execute attacks. NIST's earlier independent evaluation also found strong cyber capabilities, with differing access and safeguard conditions across models.

Analysis The new event is Anthropic's evaluation, not GLM-5.3's older release. Capability comparisons are narrow benchmark results, and a closed-model vendor has a commercial interest in the open-weight safety debate.

Anthropic / NIST / The Decoder · 29–30 Sep 2026 · Primary research · Independent government evaluation · Independent context

6. Coding-agent screenshot workarounds exposed private development images

Reporting Glow says it found over 13,000 internal images from more than 300 organizations on public GitHub repositories, often under developers' personal accounts. Agents seeking review evidence sometimes published screenshots externally; gitshot was involved in about a third of affected organizations. The Hacker News inspected that tool's public-repository behavior and notes GitHub's CLI now supports private-repository attachments.

Analysis This is a distinct multi-company disclosure, not the earlier OpenAI 53-image incident. The total comes from a security vendor whose counting method is not public; third-party downloads or resulting harm are not established.

Glow / The Hacker News · 29–30 Sep 2026 · Primary findings · Independent checks

7. Six AI companies sign a voluntary White House audit accord

Reporting Google, Meta, Anthropic, SpaceXAI, Nvidia and OpenAI signed an accord calling for internal monitoring, empowered oversight teams, independent external evaluation and board oversight. CoinDesk reports no penalties, disclosure requirement or implementation deadline. A separate executive order directs federal terminology to change from AI to 'Super Intelligence'; it does not demonstrate a technical capability threshold.

Analysis The accord is a voluntary commitment, not mandatory safety law. Its value depends on auditor independence and follow-through. The terminology order and substantive oversight pledge are separate instruments, grouped here as one White House event.

Accord text / White House / CoinDesk · 29–30 Sep 2026 · Accord text · Executive order · Independent report

8. DeepSeek open-sources programming infrastructure for Huawei Ascend

Reporting DeepSeek said it partnered with Huawei to develop and open-source Ascend programming infrastructure, including TileLang and compute and communication libraries, Reuters reports from its official WeChat statement. DeepSeek describes TileLang as a simpler high-level alternative to CUDA programming while retaining access to hardware performance.

Analysis Software portability is a major constraint on alternatives to Nvidia. This is a tooling release, not proof of CUDA-equivalent ecosystem maturity, workload performance or a new chip-purchase approval.

Reuters · 30 Sep 2026 · Report

9. General Intuition raises $220 million for game-data-trained world models

Reporting General Intuition raised $220 million at a $6.2 billion valuation, GamesBeat and SiliconANGLE report, with investors including Valor, Atreides, Seven Seven Six, Point72, Khosla and General Catalyst. The Medal-linked company uses action-labeled game video to develop world models and is opening early partner access. Its MIRA research demonstration currently generates a single game's environment.

Analysis The financing backs a different route toward spatial reasoning and physical AI. Claims about real-world transfer and effectively endless simulation are company assertions, not demonstrated general robotic competence.

GamesBeat / SiliconANGLE · 29–30 Sep 2026 · Funding report · Technical context

10. OpenAI reportedly seeks a $30 billion bridge round at a $1.4 trillion valuation

Reporting Bloomberg reports early discussions for at least $30 billion at around $1.4 trillion before new capital, after OpenAI pushed back its IPO plans. Reuters' account of that reporting says terms may change, investors are driving demand and OpenAI did not immediately comment. This is a target financing, not a closed round or a public-market valuation.

Analysis The proposed bridge shows the financing demands behind frontier development. Conflicting revenue-run-rate figures in secondary coverage are omitted; they are unnecessary to establish the reported funding talks.

Bloomberg / Reuters via CNA · 29–30 Sep 2026 · Original reporting · Reuters context

Edition notes

No exact public X post URL met the evidence threshold. The DevDay model, assistant and developer stack are distinct launches; related API and Codex features are grouped to avoid padding. The Agents API beta predates this edition: Computer Use is the new addition. Dots uses shipped GPT-6 Astra, not cancelled GPT-6.1 Astra. Benchmarks and promotional claims are attributed; OCE remains a pilot and the funding talks are not a completed deal. The White House accord is voluntary, and 'Super Intelligence' terminology does not establish a measured capability level.

AI Learning Pilot

Bring agentic AI to your classroom

TechSambad's AI Learning Pilot puts agentic AI in front of your students for one module, not a semester: a live 90-minute taught session inside your timetable, an AI-built pre-class reading brief mapped to your module, a question bank with grading rubrics, a student takeaway pack, and a written outcome report. Run it once and judge it on what your students can do afterwards — Rs 2,000 + GST for the college, no obligation.

See the pilot →