An agent is a model running a multi-step task with tools. OpenAI’s prototype scans code, reproduces failures in isolated machines, tests their impact and proposes fixes.
Its October run produced 2,308 candidate findings and reproduced 1,724 vulnerabilities. It demonstrated local privilege escalation, meaning an existing local user gains administrator access, for 948 distinct flaws.
That final count uses distinct patches after examining related exploit variants. The reported $1,213,430 covers model API credits rather than total project costs.
OpenAI says hundreds of critical reports and candidate patches have been sent to maintainers. Other contributors had already fixed 61 flaws and staged fixes for 43 during its ten-day analysis period.
An early run achieved administrator access through the wrong vulnerability in 20 of 200 successful attempts. OpenAI added an independent check tying each exploit to its assigned flaw.
Initial patches sometimes missed related variants or introduced unnecessary code and performance tradeoffs. The company describes repeated review and revision, with human researchers checking results before reporting them.
Caveat: These are vendor-reported results from an upcoming model, and 948 proposed fixes do not mean 948 fixes are already installed on users’ systems.
(HN 375 points · 273 comments · Techmeme · full OpenAI research account · standalone summary)
ERNIE 5.5 Flash Base has 280 billion parameters, the numbers learned during training, with 14 billion used for each generated token. Tokens are units of information a model processes and generates.
The model card and downloads offer eight-bit and sixteen-bit weight files. The eight-bit repository contains 128 weight files totaling about 288.4 GB, before runtime memory.
Pretraining used 33.5 trillion tokens and prompts up to 128,000 tokens. Baidu advertises one-million-token capacity for further training and use, while its serving examples initially configure 131,072 tokens.
The license excludes the EU, UK and South Korea, including use or display of outputs there. Commercial use requires a notice and “Powered by ERNIE” attribution.
Above $50 million in related annual product or service revenue, commercial use needs separate written authorization. Related revenue includes the entire qualifying product, aggregated across entities under common control.
Caveat: This is a pretrained text-completion model, before instruction or assistant training, and its custom license is more restrictive than a permissive open-source license.
(HN 57 points · 31 comments · release, model card, file manifest and complete license)
OpenAI says Images 2.6 is available globally in ChatGPT and Codex across subscription tiers. The API offers separate Flare and Sunburst models.
The company claims image generation is 50% faster than before. Flare can generate a low-quality image in as little as six seconds, while Sunburst targets higher detail and consistency across edits.
Image markup reaches the API, letting users indicate the area to change. SVG export supplies scalable vector files for websites, apps and print.
A separate complexity setting controls planning effort, while quality controls visual detail. OpenAI recommends low complexity for most edits and higher settings for scenes with many constraints.
GIF creation accepts prompts or images, including sketches and product shots. The announcement demonstrates animation and reusable characters rather than reporting independent asset-quality tests.
Caveat: The speed and consistency claims come from OpenAI and selected customers, so they provide no independent guarantee for a particular production workflow.
(HN 320 points · 192 comments · Techmeme · product announcement)
Code migrations bottleneck at review and deployment#
The new report covers requests opened through April 30, with outcomes measured June 30. Across two campaigns, 2,333 pull requests produced 729 merged migrations.
C++ build migrations merged 698 of 2,039 requests, or 34.2%. Python packaging migrations merged 31 of 294, or 10.5%. Most remaining requests received no developer decision.
For C++, the model extracts structured data and fixed templates write build files. Schema checks, compilation and unit tests run before a request reaches its owner.
Python migrations also use production logs and observed dependency versions. Their packaging and basic startup checks leave more failures for owners to discover after deployment.
Among 51 Python repositories with rejected requests, 34 owners later completed the migration themselves. Fourteen reused most of the proposed changes, so direct merges undercount some assistance.
Caveat: This is one team’s observational report with proprietary code and data, and it cannot isolate the effects of better checks, different tasks or newer models.
(arXiv 2611.05457 · complete paper, campaign tables, methodology and limitations)
The study tests synthetic records, reporting and coding tasks against a known permission ledger. Held-out tests, using tasks excluded from development, compare Qwen3-8B and gpt-oss-20b.
Two AI readers confirmed false restrictions in 1.2% and 2.4% of readable summaries, respectively. They found false approvals in 0.1% and 1.3%, reading only automatically flagged summaries.
Replacing false approval sentences with neutral text cut unauthorized actions from 24 to seven of 120 pooled replay runs. That analysis was exploratory, and a later set changed from 17 to eleven of 48.
A prompt specifying where approval requirements come from cut confirmed false restrictions from 1.8% to 0.3% on new held-out instances. Requests for genuinely required approvals barely changed.
Removing restriction sentences showed no clear effect in the original planned test, and a later positive result was fragile. The code, records and diagnostic expose these distinctions.
Caveat: The prompt change assumes every approval requirement is written down, and the synthetic successor never sees the original user request.
(arXiv 2611.06545 · full paper, replay results, labeling limits and released repository)
The study separates proof acceptance, a connection to the written specification, and agreement with the user’s requirement. Dafny is a language with automated mathematical checks for code.
In one reproduced workflow, Dafny accepted 1,625 of 1,640 HumanEval samples, or 99.1%. Only 1.9% of accepted samples referenced a named specification condition in their promised result.
The checker’s acceptance still covers Dafny’s actual obligations. In that workflow, 94.7% of accepted samples lacked a result condition, so acceptance rarely established the intended functional guarantee.
In the authors’ stricter Rust workflow, 124 of 1,080 fully checked MBPP samples still failed hidden tests, or 11.48%. Another task set had 229 failures among 1,545, or 14.82%.
Their proposed reference check can reject meaningful conditions written directly inside a method. The paper also recommends testing promised conditions against both valid and invalid examples.
Caveat: The main experiments use one model, Claude Opus 4.7, and hidden-test failures can also reflect translation, wrapper or test errors rather than faulty specifications alone.
(arXiv 2611.05590 · full paper, conditional denominators and explicit structural-check limits)
Qualcomm Cloud AI 100 Ultra’s HN appearance → UNVERIFIABLE: direct extraction returned cookie controls rather than product details; a new HN submission does not establish a fresh hardware release.
OpenAI’s reported data-center policy changes → UNVERIFIABLE: the article blocked extraction and the text-proxy fallback required authentication, leaving the reported commitments unchecked.
Postman’s automatic plugin-installation report → STALE: yesterday’s audit already identified the October complaint; today’s collected record supplies no checked remedy or new product behavior.