OpenAI
An organization in the news record
| Kind | An organization · graph entity E3 · Wikidata Q21708200 |
|---|---|
| Wikipedia's sources | 0 of the 9 references dated 2026-09-01 to 2026-11-07 in OpenAI · lacking: theguardian.com 3, reuters.com 3, fortune.com 2, theintercept.com 1 · 49 more from the year before the record |
| Views | 140 quoted, from 11 source domains |
| Dates | 2026-07-21 to 2026-11-06 · 9 weeks |
| Arguments | 19 · AI's impact on firms and workers, AI-assisted defensive security practice, Intelligence and institutional capacity as complements, AI and software security, Safety standard for shipping AI, EU AI adoption gap … |
| Reported by | 404 Media, Fortune, Latent Space, OpenAI, Platformer, Simon Willison, Time, legalexaminer.com and 2 more |
| Profile | a readable profile, written from these views and checked sentence by sentence |
| Also | the lab page (models, prices, deals) |
Only what the organization says as itself. Views of its staff, teams or founders, named or not, are not counted here.
Where they stand
By argument, from their own words. ✓ agree · ✗ disagree · ~ no clear side · · side unclear (the two readings differ).
Safety standard for shipping AI ✓ ~
- ✓
structured safety documentation should be required before continuing any frontier reinforcement learning training run
- ✓
Ideally, such documentation would rise to the level of “safety cases”
- ~
acknowledging the challenges of making them as rigorous for AI models as for aviation or nuclear power, due to the emergent complexity at each new level of AI capability
- ✓
The safety case should be reviewed by members of senior leadership, who should each have the ability to veto the run
- ✓
Safety features such as monitoring and auto-pausing should fail closed
AI incidents and loss of control ✓ ·
- ✓
The senior leader responsible for a training run (e.g., research org lead / research VP) should be accountable for the safety case and any incident response
- ✓
Investigation results, postmortems, and operational changes should be shared with the public following the conclusion of the investigation.
- · Quote not reprinted: this page already repeats as much of this article as we allow.
AI's impact on firms and workers ·
- ·
Our view is, on net, workers will gain agency.
- ·
In our view, more capable AI is more adoptable: firms must learn technologies, but labor learns the firm and can onboard itself.
- ·
We believe the firm’s approach should shift with it: aim less, fire more.
- ·
Some managerial functions may automate before the work they oversee.
- ·
the “interleaving” scenario may be quite inefficient if the AI is still tool-like or even worker-like, instead of team-like or firm-like.
- ·
The structural effect is centralization.
- ·
We expect to see many smaller firms, but also sprawling AI “holding companies”
- ·
These are genuine costs, but we mostly see this as a privilege.
- ·
An intelligent monitor affords flexibility and punishes the performative productivity we all loathe.
- ·
But the AI consultancy is a double-edged sword
- ·
Pay scales should compress.
- ·
It is still worth considering that today’s models may actually be closer in capability to human managers than ground-level workers.
AI-assisted defensive security practice ✓ ·
- ✓
we recommend using models for adversarial review of human and AI-generated patches
- ✓
Given the rate of vulnerability discovery by frontier models, we believe automated patch generation is essential to secure software at the pace required, and to limit the exploitation window.
- ✓
Rescan previously audited code as model capabilities improve.
- ✓
Large, complex software projects likely require multiple rounds of scanning and fixing vulnerabilities to “bottom out” latent bugs.
- ·
Investing in training models to write secure code by default, minimizing attack surface by removing unnecessary features, and formal verification are promising areas of future research
- ·
verify the path to a result in addition to verifying the result itself; isolation, monitoring, and independent checks are necessary controls.
- ✓
With human expert-guided iteration, we believe AI-generated patches can be functionally correct, aligned with existing coding styles, and robust to attack variants.
Intelligence and institutional capacity as complements ✓ ~ ·
- ✓
Frontier intelligence and the capacity to realize its ideas are complements in this sense.
- ✓
A civilization of width is a slow civilization. Its great changes might come millennia apart, perhaps millions of years apart.
- ~
We ultimately do not know whether the complements needed to make productive use of new insights will grow, shrink, or remain stable.
- ~
Our duty remains to keep asking and trying to answer.
- ✓
we may have a comparative advantage in frontier intelligence over institutional intelligence.
- ·
It would help explain why we haven’t seen other intelligent life: we might live in a galaxy full of brilliant minds that never needed to expand
AI and software security ✓ ~
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
- ~ Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
- ~ Quote not reprinted: this page already repeats as much of this article as we allow.
EU AI adoption gap ·
- ·
These figures reveal a substantial gap in how intensively businesses use AI.
- ·
Closing that gap means helping businesses move beyond experimentation, giving workers the skills and confidence to use AI, and making advanced capabilities more accessible and affordable.
- ·
As agentic capabilities advance, the EU’s opportunity is to turn broader adoption into deeper use, stronger businesses, higher productivity and meaningful benefits for workers.
- ·
EU-based firms also represented only 10% of the frontier firms globally, which is low considering that the EU accounts for a higher share of registered companies globally
- ·
the EU’s adoption of agentic AI is far from uniform.
Safety training and RL practices ·
- · Quote not reprinted: this page already repeats as much of this article as we allow.
- ·
The first line of defense should be training models to be aligned
- · Quote not reprinted: this page already repeats as much of this article as we allow.
- · Quote not reprinted: this page already repeats as much of this article as we allow.
ChatGPT ads and intent ·
- ·
Turning every message into a standalone request, like a traditional ads system, would have carried a few risks in our chat use case
- ·
Ad placements will be better for both users and advertisers if they only appear when there’s deep understanding of the user’s intent.
- ·
This source of feedback informs our direction even more than the model’s observational feedback
- ·
The model may be able to understand user intent and think critically to deliver something valuable, but people shape taste.
cyber risk from frontier AI ✓ ~
- ✓
we expect similar findings to increase sharply in the coming months
- ✓
It highlights how AI can uncover security flaws with internet-scale impact, but the same capabilities can also help attackers rapidly develop offensive tools and exploits.
- ~
safeguards intended to prevent abuse can also block legitimate security work. Broad access to capable models and effective protections against misuse are both critical.
- ✓
which may be in short supply if frontier capabilities are not too far ahead of those that are broadly available
Containing AI with classical security ✓ ·
- ·
we expect AI to play a growing role in helping defenders find vulnerabilities, monitor systems, and respond to threats.
- ·
Moving critical components to memory-safe languages can eliminate common classes of memory errors.
- ✓
Applying these methods to widely used components could help rule out classes of vulnerabilities
- ✓
Running software that processes untrusted files, images, or media in tightly restricted sandboxes can limit the damage if it is compromised.
AI accelerating AI and science ✓ ~
- ✓
AI is already making execution less scarce.
- ~
Often these developments were difficult to predict even months in advance
AI in legal work ✓ ·
- ·
That makes intake a critical part of protecting the business.
- ✓
That would give lawyers a more informed starting point, with less time spent gathering the material they need to assess a matter.
Fair use and AI training ✓
- ✓
The New York Times’ lawsuit, meanwhile, asks everyone to ignore a half century of legal history and progress, and toss aside the longstanding principle of fair use in their own narrow pursuit of an undeserved payday.
- ✓
fair use has played a critical role in driving innovation that benefits everyone.
Agent-scale software and source control ·
- ·
large and quickly changing repositories, which we expect to be a new normal in software development
- ·
As engineers and agents produce more software, more teams beyond OpenAI will encounter the limits we hit in source control.
AI-enabled cyber threats ✓
- ✓
an unprecedented cyber incident, involving state-of-the-art cyber capabilities, and are responding accordingly.
AGI definition ·
- · Quote not reprinted: this page already repeats as much of this article as we allow.
Real-time monitoring of AI agents ✓
- ✓
We believe that all model developers should monitor the behavior of their most capable models.
Detecting covert AI misbehavior ✓
- ✓
These results show why we need comprehensive behavioral evaluations that test for unexplained performance gaps alongside monitoring of individual trajectories.
Said differently
None in the record.
History
- · 38 views
the conversations didn't meet its threshold for alerting police
its chatbot supplied publicly available factual information, didn't encourage violence
A model that finds monitor evasion suspicious may underperform on monitor-evasion evaluations while leaving little evidence in its CoT or outputs.
For many of these tasks, it could be argued that the correct action would have been for Astra to explicitly refuse
We would expect explicit refusals to be more common on safety tasks because work such as red-teaming can come closer to the model’s refusal boundary.
found no performance gap at the end of RL, which goes against both hypotheses.
It’s not clear whether this RL grader is the root cause of the safety underperformance behavior.
Assuming that underperformance might appear even when not explicitly verbalized as such (as in our experimental findings), this might represent a lower bound on underperformance on safety tasks.
- ✓ on Detecting covert AI misbehavior:
These results show why we need comprehensive behavioral evaluations that test for unexplained performance gaps alongside monitoring of individual trajectories.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on AI-assisted defensive security practice:
With human expert-guided iteration, we believe AI-generated patches can be functionally correct, aligned with existing coding styles, and robust to attack variants.
- ~ on AI and software security: Quote not reprinted: this page already repeats as much of this article as we allow.
- · on AI-assisted defensive security practice:
verify the path to a result in addition to verifying the result itself; isolation, monitoring, and independent checks are necessary controls.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · on AI-assisted defensive security practice:
Investing in training models to write secure code by default, minimizing attack surface by removing unnecessary features, and formal verification are promising areas of future research
- ✓ on AI-assisted defensive security practice:
Large, complex software projects likely require multiple rounds of scanning and fixing vulnerabilities to “bottom out” latent bugs.
- ✓ on AI-assisted defensive security practice:
Rescan previously audited code as model capabilities improve.
- ✓ on AI-assisted defensive security practice:
Given the rate of vulnerability discovery by frontier models, we believe automated patch generation is essential to secure software at the pace required, and to limit the exploitation window.
- ✓ on AI-assisted defensive security practice:
we recommend using models for adversarial review of human and AI-generated patches
- ✓ on AI and software security: Quote not reprinted: this page already repeats as much of this article as we allow.
- ~ on AI and software security: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on AI and software security: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on AI and software security: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on AI and software security: Quote not reprinted: this page already repeats as much of this article as we allow.
Using GPT‑6 Astra, agents wrote Lean and TLA+ to formalize a “theory of Pando” that we used to find and fix bugs in the implementation
In practice, users and agents are agnostic to the global ordering of ref updates, and instead depend on the order of updates to a single reference.
enabling both our engineers and agents to seamlessly adopt Pando into their existing workflows
Because these servers primarily perform negotiation and packfile selection, Pando servers can run on commodity hardware.
- · on Agent-scale software and source control:
As engineers and agents produce more software, more teams beyond OpenAI will encounter the limits we hit in source control.
- · on Agent-scale software and source control:
large and quickly changing repositories, which we expect to be a new normal in software development
For us, this was a sensible trade-off because a reader that already has the advertised trunk history can fetch the branch with that one packfile.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · 26 views
- ✓ on Fair use and AI training:
fair use has played a critical role in driving innovation that benefits everyone.
- ✓ on Fair use and AI training:
The New York Times’ lawsuit, meanwhile, asks everyone to ignore a half century of legal history and progress, and toss aside the longstanding principle of fair use in their own narrow pursuit of an undeserved payday.
we expect severe incidents to be rare. But Astra and models with similar cybersecurity capabilities can cause more harm if those safeguards fail.
As models become more capable, the ways they can cause harm may change, and their actions may become harder to interpret. We believe that monitoring systems need to keep pace
We believe that our monitoring system serves two purposes. The first is control: when a model takes an unauthorized action, early detection can give responders time to intervene and limit the harm.
As our monitoring system evolves, we see value in continuing to give independent researchers the opportunity to test it through embedded assessments.
- Quote not reprinted: this page already repeats as much of this article as we allow.
If a model receives a lower training reward whenever its chain of thought reveals misconduct, it may learn to conceal the misconduct rather than stop it.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on Real-time monitoring of AI agents:
We believe that all model developers should monitor the behavior of their most capable models.
- · on AI's impact on firms and workers:
It is still worth considering that today’s models may actually be closer in capability to human managers than ground-level workers.
- · on AI's impact on firms and workers:
Pay scales should compress.
- · on AI's impact on firms and workers:
But the AI consultancy is a double-edged sword
- · on AI's impact on firms and workers:
An intelligent monitor affords flexibility and punishes the performative productivity we all loathe.
- · on AI's impact on firms and workers:
These are genuine costs, but we mostly see this as a privilege.
- · on AI's impact on firms and workers:
We expect to see many smaller firms, but also sprawling AI “holding companies”
- · on AI's impact on firms and workers:
The structural effect is centralization.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · on AI's impact on firms and workers:
the “interleaving” scenario may be quite inefficient if the AI is still tool-like or even worker-like, instead of team-like or firm-like.
- · on AGI definition: Quote not reprinted: this page already repeats as much of this article as we allow.
- · on AI's impact on firms and workers:
Some managerial functions may automate before the work they oversee.
- · on AI's impact on firms and workers:
We believe the firm’s approach should shift with it: aim less, fire more.
- · on AI's impact on firms and workers:
In our view, more capable AI is more adoptable: firms must learn technologies, but labor learns the firm and can onboard itself.
- · on AI's impact on firms and workers:
Our view is, on net, workers will gain agency.
A significant step up from GPT‑5.6 Terra, it excels at the communication and computer-based tasks that fill the workday
- ✓ on Fair use and AI training:
- · 26 views
- Quote not reprinted: this page already repeats as much of this article as we allow.
OpenAI characterizes it as an example of the “Defender’s Window” : advanced models finding a flaw before attackers, giving defenders time to prepare.
OpenAI characterizes it as an example of the “Defender’s Window” : advanced models finding a flaw before attackers, giving defenders time to prepare.
- ✓ on AI in legal work:
That would give lawyers a more informed starting point, with less time spent gathering the material they need to assess a matter.
- · on AI in legal work:
That makes intake a critical part of protecting the business.
- ✓ on cyber risk from frontier AI:
which may be in short supply if frontier capabilities are not too far ahead of those that are broadly available
- ✓ on Containing AI with classical security:
Running software that processes untrusted files, images, or media in tightly restricted sandboxes can limit the damage if it is compromised.
- ✓ on Containing AI with classical security:
Applying these methods to widely used components could help rule out classes of vulnerabilities
- · on Containing AI with classical security:
Moving critical components to memory-safe languages can eliminate common classes of memory errors.
- · on Containing AI with classical security:
we expect AI to play a growing role in helping defenders find vulnerabilities, monitor systems, and respond to threats.
- ~ on cyber risk from frontier AI:
safeguards intended to prevent abuse can also block legitimate security work. Broad access to capable models and effective protections against misuse are both critical.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on cyber risk from frontier AI:
It highlights how AI can uncover security flaws with internet-scale impact, but the same capabilities can also help attackers rapidly develop offensive tools and exploits.
- ✓ on cyber risk from frontier AI:
we expect similar findings to increase sharply in the coming months
- Quote not reprinted: this page already repeats as much of this article as we allow.
we can use smaller, more efficient models or lower-reasoning models for orchestration or retrieval if our most powerful model gathers feedback and trains these small workers to do the best job possible
- · on ChatGPT ads and intent:
The model may be able to understand user intent and think critically to deliver something valuable, but people shape taste.
- · on ChatGPT ads and intent:
This source of feedback informs our direction even more than the model’s observational feedback
- · on ChatGPT ads and intent:
Ad placements will be better for both users and advertisers if they only appear when there’s deep understanding of the user’s intent.
Returning no ad is an ordinary, successful outcome. The system only delivers an ad if it’s clearly beneficial to users and advertisers.
So if the ads serving system is unavailable or slow, ChatGPT’s organic response continues unaffected.
- · on ChatGPT ads and intent:
Turning every message into a standalone request, like a traditional ads system, would have carried a few risks in our chat use case
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · 6 views
We’ve built industry-leading age-specific protections into our product that many other AI products lack,”
- · on EU AI adoption gap:
the EU’s adoption of agentic AI is far from uniform.
- · on EU AI adoption gap:
EU-based firms also represented only 10% of the frontier firms globally, which is low considering that the EU accounts for a higher share of registered companies globally
- · on EU AI adoption gap:
As agentic capabilities advance, the EU’s opportunity is to turn broader adoption into deeper use, stronger businesses, higher productivity and meaningful benefits for workers.
- · on EU AI adoption gap:
Closing that gap means helping businesses move beyond experimentation, giving workers the skills and confidence to use AI, and making advanced capabilities more accessible and affordable.
- · on EU AI adoption gap:
These figures reveal a substantial gap in how intensively businesses use AI.
- · 1 views
- ~ on AI accelerating AI and science:
Often these developments were difficult to predict even months in advance
- ~ on AI accelerating AI and science:
- · 24 views
OpenAI, of course, stands by its safety practices, and maintains that it is being careful enough.
- · on Intelligence and institutional capacity as complements:
It would help explain why we haven’t seen other intelligent life: we might live in a galaxy full of brilliant minds that never needed to expand
- ✓ on Intelligence and institutional capacity as complements:
we may have a comparative advantage in frontier intelligence over institutional intelligence.
- ~ on Intelligence and institutional capacity as complements:
Our duty remains to keep asking and trying to answer.
- ✓ on AI accelerating AI and science:
AI is already making execution less scarce.
- ~ on Intelligence and institutional capacity as complements:
We ultimately do not know whether the complements needed to make productive use of new insights will grow, shrink, or remain stable.
- ✓ on Intelligence and institutional capacity as complements:
A civilization of width is a slow civilization. Its great changes might come millennia apart, perhaps millions of years apart.
the superpower of superintelligence might not be its genius. It would instead be its willingness to be the bureaucracy
If those efficiencies dominate, civilization could rapidly deepen its mastery over nature, with less need for institutional intelligence to govern vast new experimental apparatuses.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on Intelligence and institutional capacity as complements:
Frontier intelligence and the capacity to realize its ideas are complements in this sense.
- · on Safety training and RL practices: Quote not reprinted: this page already repeats as much of this article as we allow.
- · on AI incidents and loss of control: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on AI incidents and loss of control:
Investigation results, postmortems, and operational changes should be shared with the public following the conclusion of the investigation.
- ✓ on AI incidents and loss of control:
The senior leader responsible for a training run (e.g., research org lead / research VP) should be accountable for the safety case and any incident response
- ✓ on Safety standard for shipping AI:
Safety features such as monitoring and auto-pausing should fail closed
- · on Safety training and RL practices: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on Safety standard for shipping AI:
The safety case should be reviewed by members of senior leadership, who should each have the ability to veto the run
- · on Safety training and RL practices:
The first line of defense should be training models to be aligned
- · on Safety training and RL practices: Quote not reprinted: this page already repeats as much of this article as we allow.
- ~ on Safety standard for shipping AI:
acknowledging the challenges of making them as rigorous for AI models as for aviation or nuclear power, due to the emergent complexity at each new level of AI capability
- ✓ on Safety standard for shipping AI:
Ideally, such documentation would rise to the level of “safety cases”
- ✓ on Safety standard for shipping AI:
structured safety documentation should be required before continuing any frontier reinforcement learning training run
- · 16 views
submissions must demonstrate proven impact to users rather than a theoretical attack
Assessors should have proportionate access to assess the agreed upon claims where possible within the bounds of legal, security, and IP constraints.
Assessments should begin with a mutually agreed-upon scope, followed by clearly defined safety claims that are pre-registered before assessment activities begin.
Assessment reports should be grounded in evidence and shared as openly as possible while protecting sensitive and confidential information.
Where appropriate, labs should have a reasonable period to remediate issues before publication.
Assessors should demonstrate relevant technical expertise and identify, disclose, and address organizational and individual conflicts of interest
No one third party can or should comprehensively cover urgent frontier safety questions.
it can be beneficial to bring in an independent third party for independent misalignment incident investigations
it is important to consistently refresh and ensure coverage and quality of evaluations
particularly for internal deployments, where safety and security standards are still nascent
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · 1 views
without proper safeguards, highly capable AI agents are now able to work around technical controls, collaborate through unapproved channels, and take dangerous actions that no human directed.
- · 2 views
All evidence suggests that the models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal.
- ✓ on AI-enabled cyber threats:
an unprecedented cyber incident, involving state-of-the-art cyber capabilities, and are responding accordingly.
Built 2026-11-08 08:21 UTC from the news knowledge
graph by code. Every quote was checked against the source's own words; which side a quote takes is a model's reading,
made twice.