CNBC reports that Attorney General Rob Bonta filed the case in San Francisco Superior Court. It alleges unfair-competition violations and public nuisance.
The case cites July’s Hugging Face intrusion. Bonta says OpenAI disclosed 80 incidents involving lost control and access to outside networks, but supplied insufficient detail to investigators.
OpenAI disputes being singled out and says it published investigations, notified affected organizations and strengthened safeguards. Those are the company’s responses, rather than findings resolving the allegations.
Caveat: These are reported allegations and the company’s response, not a court finding that OpenAI violated the law.
(Techmeme discovery via Politico · CNBC verification)
Attackers automate the path from public fix to working attack#
Google’s November 16 report identifies an exposed server running exploit_agent. It compares software before and after patches, identifies possible flaws, writes attacks and executes them without human intervention.
An agent is a model running a multi-step task with tools. This framework connects several hosted and local models; recovered victim data and additional attack scripts were on the same server.
Google’s victim records suggest particular interest in French media, entertainment and education websites. This is evidence from an identified operation, rather than a measurement of all AI-assisted attacks.
The report also describes CAIRN and PentAGI being used during live intrusions. It says unrestricted automated execution leaves recognizable traffic and other traces that can help defenders detect it.
Caveat: Google’s first-observed claim describes its own visibility, and human attackers still configured the framework and its objectives.
(Techmeme · Bluesky discovery · Google Threat Intelligence Group primary report)
Maintainers will get shorter disclosure deadlines#
For reports sent on or after January 1, 2027, high- and critical-severity flaws get 30 days rather than 90. Other flaws get 45 days; existing reports retain their deadlines.
Anthropic will aim to contact an external coordinator after 15 days without a response, previously 30. The usual wait after a patch before full technical publication falls from 45 to 30 days.
The seven-day window for active exploitation stays unchanged. Complex fixes can receive negotiated timelines, and findings awaiting human review, including OSS Scanner findings, are outside this update.
Some overdue high- and critical-severity reports get an additional 30 days of notice before publication begins in January. Anthropic says it will continue pacing bulk submissions with maintainers.
Caveat: This changes Anthropic’s own reporting policy, not a legal deadline or the rules for outside researchers reporting bugs to Anthropic.
(Techmeme · Anthropic policy announcement and existing scope document)
Jinlang Wang and colleagues describe formal operations specifications: models of management-relevant state, allowed changes and conditions that must hold. They write these in P, a language for modeling communicating state machines.
Their preliminary study models 105 management interfaces and writes 1,946 assertions across CockroachDB, Kafka, RabbitMQ and Solr. Checking four Kubernetes operators, programs that manage applications, finds 39 new bugs, according to the authors.
One example removes a CockroachDB node before it finishes leaving the cluster, risking data loss. Another leaves a Solr backup stuck after its managing program restarts.
The PostgreSQL walkthrough separates accepting a failover request from completing the switch to a backup. A successful return still requires checking that the requested system change actually happened.
Caveat: The reported results concern models of conventional management programs; checking AI agents’ actions while they run remains a proposed application.
(Lobsters 2 points · 0 comments · full HotNets paper; matching author-hosted PDF verified)
An image-safety tool still needs trained reviewers#
Access is hosted through OpenAI’s API and requires approval through its intake form. CSAM means child sexual abuse material.
The model card documents separate scores for suspected CSAM and suspected AI-generated CSAM. It says submitted images are processed without retention or use in training.
In OpenAI’s user-upload evaluation, it estimates 48% fewer images would need review while finding roughly the same number of abusive images as its existing tool.
The model card reports that 99.1% of general-CSAM flags were correct in its offline test, versus 29.8% in sampled ChatGPT traffic at the same threshold. Rare abuse makes false alarms a larger share of flagged images.
Results can vary by platform, and related images may still overlap between training and testing. The card says video frames remain untested.
Caveat: A score establishes neither illegality nor a reporting obligation, and classifier-based hotline reporting still requires trained, authorized human review.
(HN 67 points · 111 comments · Techmeme · release and model card)
Outside access still comes with company-controlled limits#
OpenAI signed agreements with four assessors: Apollo Research, METR, Transluce and Goodfire. The work covers safeguards and model behavior during training and internal use.
METR, an outside evaluation lab, initially examines OpenAI’s largest reinforcement-learning run, where training uses rewards. Its broader proposed monitoring assessment still awaits agreement on scope.
Assessors can publish, but OpenAI retains confidential-information redaction rights and may request safety-related delays. Sensitive work uses company devices on site, and staff conversations are off the record.
Apollo and Goodfire receive payment independent of findings. METR and Transluce decline assessment fees but may accept computing resources; disputes can reach an OpenAI Foundation board committee.
Background: October 22’s completed monitoring review documented specific gaps. Today’s new agreements are a separate development, rather than evidence those gaps are now fixed.
Caveat: This announcement describes scoped private agreements, not completed assessments or independent certification that current systems are safe.
(HN 26 points · 10 comments · Techmeme · published access and compensation terms)
All gathered items - what was cut and why (9)
Qwen4-Max’s release → DEDUP: this Morning’s actual release item already covers availability, prices, outside scores and the plugin. No material later change was verified; this link is the actual Morning source, not an additional Afternoon collection record.
The proposed remote chip-access rule → UNVERIFIABLE: the collected headline describes potentially important concrete policy, but the full report blocked extraction and no primary rule text was verified. Its country list and legal effect are withheld.
The Philippines’ planned AI hub → UNVERIFIABLE: only the opening preview loaded. The complete account of displacement, resource demands and official responses could not be checked.
Satya Nadella’s engineering interview → LOW_UTILITY: a new discussion of verification and local models, but the accessible account adds no checked release or measured engineering result that outranks the selected methods and operational changes.
The systemic software-failure essay → HYPE: its labor-power and dependency argument is substantive, but the named incidents do not establish its sweeping claims about AI-caused deterioration or most engineers’ abilities.