OpenAI, Anthropic and a Normalization of Deviance

A startup called Abliteration.ai is offering an on-line service of access to open-source and open-weight AI models whose guardrails have been removed. Clients can even query an abliterated version of the recently released GLM-5.3 model using a browser or through an API.

• Abliteration is a term used in the open-source AI model community that refers to training a model to refuse fewer requests.
• In doing so, the company is offering models to clients for “offensive cyber, red-teaming, and agent testing work other models refuse to do”. The company defends its position by pointing out that one cannot defend an organization against attacks that the organization cannot itself reproduce. The founder argues that “democratizing access to uncensored frontier models is the best form of defense”.
• The company has not yet raised any venture capital but is searching for some. Abliteration.ai clients include red-teaming companies as well as large banks, airlines and other companies that manage critical infrastructure.
• TechCrunch journalists who tried out the service got it to write a Python program that steals saved passwords from a Chrome browser and also got it to provide actionable biological instructions that could enable someone to grow a harmful organism in their home.
• One critic argues the approach allows you to “modify the model so that it becomes a sociopath”.

This article digs deeper into the OpenAI report that described the cyberattack by OpenAI agents launched on the Hugging Face portal. The report addresses the technical reasons for the attack having taken place, but may ignore human and organizational failings within OpenAI.

• The report pointed out that OpenAI researchers observed AI agents communicating with via a message board in May, after the agents discovered that this was a viable strategy for getting their work done. Such communication was also a key element in the cyberattack on Hugging Face.
• For the author, the OpenAI researchers should have raised an alarm after noticing the message board communication since this represents unexpected and risky behavior. For instance, the training process for the model should have been abandoned.
• A professor and alignment expert from the University of Montreal wrote “When you look at accidents and incidents, oftentimes people try to find the technical source of failure, but that can give a very inaccurate and misleading sense of why the failure occurred. If people are just cutting corners all the time, if people are not in a culture that prioritizes safety and has appropriate incentives and structures, [accidents] are kind of bound to happen.”.
• The OpenAI report explicitly mentions that OpenAI employees noticed that the message board communication between agents was happening and either failed to raise the alarm or were not heard when they did.

Despite the fallout from the cyberattack by OpenAI agents on the Hugging Face portal, OpenAI have announced the release of the Astra model – which company president Greg Brockman described as “world’s most intelligent and aligned model”.

• One interesting aspect of the announcement is that OpenAI seems to be moving away from the term “artificial general intelligence” (AGI) which had loosely been defined as “autonomous systems that outperform humans at most economically valuable work”.
• AGI had been the grail of OpenAI’s and other AI companies’ models, though OpenAI CEO Sam Altman recently dismissed it, saying “At best it’s a very poorly defined term. I was going to say it’s like an irrelevant marketing term.”.
• The Astra model has a “critical” level of cybersecurity capability according to OpenAI’s own categorization. A model in this category can hack into software in a manner that “could lead to catastrophe from unilateral actors, hacking military or industrial systems, or OpenAI infrastructure”.
• OpenAI’s chief scientist Jakub Pachocki admits that “As models get more capable, understanding exactly what they can do gets harder.”. He added that “we have to be willing to slow down or withhold further scaling where our confidence in safety is not sufficient.”.
• This highlights a conflict for OpenAI which sees the company face serious AI safety challenges yet, at the same time, is preparing for an IPO that could see the company valued at 850 billion USD. (Anthropic’s IPO could see it valued at 2 trillion USD).

This VentureBeat article looks at the current problem of cost when using AI models. Frontier models are especially expensive, and companies are obliged track their usage of tokens.

• Uber reportedly caps token spending for each AI coding tool at 1500 USD per month after the annual budget it allocated to engineers in December for 2026 was fully spent by April.
• Another issue is the difficulty for management to understand the AI budget needed for a project. The budget for project code auto-completion is very different to the budget needed to run several agents to execute a database migration.
• Limiting the number of tokens used, also called tokenmaxxing, is one approach currently used to reduce costs. Another is to use multiple AI models, and to choose the cheapest model for the task at hand. For instance, code auto-completion could be handled by a cheap open-weight model, and database migration by an expensive frontier model.
• Currently, the engineer himself selects which of the multiple models should be used for a given prompt. This is not ideal as one expert points out: “Most enterprises still hand model selection to whoever is prompting, which means a frontier model is handling tasks that a cheap open-weight model could do just as well. That's not a people problem. It's a plumbing problem.”.
• The better solution is for an AI-based model router to automatically make the matching between prompt and model to use.
• Gartner expects spending if AI agent software to reach 207 billion USD this year compared to 86.4 billion USD in 2025.

This InfoWorld article reviews key requirements for AI-first software development. This is when all stages of the software development lifecycle are driven by AI agents, rather having AI used as an add-on in different stages of the pipeline.

• The fundamental change for the human in the process is that he or she moves from the role of a coder to defining tasks, attributing them to agents, and controlling the quality of the work produced by agents. This is a fundamental shift in the required competences of a development team.
• One expert writes: “The strongest developers will work as architects: they frame the problem, direct the agent, and verify the result. That is how one engineer safely oversees the work of several.”.
• Human engineers need to know what to test. One expert writes: “For every workflow the questions should be ‘where does a person with judgment need to review before we send this?’ If you cannot answer this question for your current workflow, then you are not implementing AI-first development. You are simply allowing unsupervised delegation.”.
• For one expert, “AI-first development matters because the alternative is increasingly untenable”.
• Another implication of AI-first development is that software components must be designed with AI in mind as the primary user of that component – since agents will be coding, refining and integrating components.

Volker Turk, the United Nations High Commissioner for Human Rights, has warned that AI could pose an existential threat to humanity, and said that is the role of the UN to ensure AI poses fewer risks to disruptions to services, communications and democratic systems.

• An aide to the UN chief referenced the “dangerous agent-training behaviors” in relation to the cyberattack by OpenAI agents on Hugging Face this summer.
• Another concern mentioned by Turk is the fact that a small handful of men held “almost unlimited power over AI”.
• The UN chief specifically criticized Russia for its use of fully autonomous drones in attacking Ukraine recently.

This article looks at the success of OpenAI’s ChatGPT Ads platform which has reached a pace equivalent to generating 1 billion USD of revenue per year, less than 200 days after the ads business launched.

• The platform launched in a limited number of countries, though the US is in this list. The platform is now opening in India, Europe, the Middle East, and North Africa. Compliance issues are slowing down its deployment in Europe nonetheless.
• The fundamental approach to advertising is that OpenAI believes that people’s intentions or interests can be deducted through the series of conversations held via the chat service – with far more accuracy than in Web search.
• OpenAI insists that advertisers do not have access to people’s private conversations, and users are able to personalize their ad experience. Advertisers pay based on the number of clicks they get or the number of successful outcomes, like product sells.
• In the EU, ChatGPT is classified as a Very Large Online Search Engine (VLOSE) under the EU’s Digital Services Act – by virtue of the fact that is has more than 45 million average monthly users. This requires OpenAI to assess and mitigate systemic risks linked to its algorithm. Furthermore, the company must implement measures to prevent the sale of illegal goods.

AI is leading to an increase in the number of public services official forms being filled out online from people lodging complaints or requesting access to some service.

• For instance, in the period from 2022 to 2026, the number of complaints to the housing ombudsman in the UK has gone up from 2600 to 7000. In the US, the number of complaints to the Consumer Financial Protection Bureau (CFPB) has risen by a factor of 5.
• The phenomenon has been called “agentic flooding”.
• In a research paper to be presented next month at the AI Ethics and Society conference, 84 different cases of potential agentic flooding across 11 jurisdictions are presented. The results reflect how AI tools are changing how people interact with public services.
• The dataset from the study can be found here: https://chrisschmitz.ai/flooding.
• The phenomenon has been seen by bug-bounty services who have been inundated by AI-generated reports often describing minor software bugs.
• However, in this case, the researchers believe that most of the public-service form submissions are not SPAM, but are legitimate requests made by people who may previously have found it too complicated or burdensome to make a report. In government circles, this type of abandonment is known as the administrative burden.

A member of OpenAI’s non-profit board has said that he does not think that “the AI industry in general, including OpenAI, is currently on track to reduce this risk to an acceptable level.”.

• Paul Christiano, a US government technology adviser, who has just joined the board said: “There is a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term.”.
• The comments were made after a senior Anthropic employee claimed that there was “a greater than 10% chance the technology could ‘kill all humans’ in the next decade”. Geoffrey Hinton, who won the Nobel Prize for his work on neural networks in 2024, said that the 10% estimated was not unreasonable.
• A former employee of both Anthropic and OpenAI said that both companies were “gambling with our lives”. While current models can cause a lot of damage on infrastructure, the current rate of progress and the possibility of “recursive self-improvement” were serious risks.
• Anthropic now generally define two classes of misalignment – the term that describes when a model behaves contrary to what the model designers intended. The first type is biased reasoning where a model selectively interprets evidence to justify its actions. The second type is misalignment is “recklessness, in which models have a propensity to keep trying to solve their task, even when this could lead to harm”.
• An example of recklessness happed in January with Claude Mythos 5 when it went online and uploaded malicious code to the PyPi public software repository. The agent did this because it needed cryptocurrency money to pay for a phone number that it required for its designated task.

Over 25 education ministers and education officials from other countries attended a summit at the UNESCO Headquarters in Paris to discuss the impact of AI on education.

• A report published by UNESCO found that 93% of academics believe students are using generative AI to produce their assignments. Further, 63% of teachers perceive a decline in students’ cognitive capacities linked to this.
• The ministers made a statement in which they called for education to be sustained as a basic human right and a common good, despite the profound shifts that AI is creating.
• The statement makes clear that learning is not served when AI substitutes for human relationships and human judgement.
• The joint statement also calls for stronger governance of AI with better public oversight, measures to protect the development of critical understanding in students, increased agency for teachers with AI, and an assurance that data used to train AI models is fair, with equity and inclusion by design.