AI Leadership Weekly

Top Stories

OpenAI finally releases GPT-OSS, but there are problems
After much delay, OpenAI has released two large-scale open-weight language models: gpt-oss-120b and gpt-oss-20b. Both models promise strong performance at relatively low cost, aiming to give researchers and developers flexible, high-quality tools to run and customise on their own infrastructure. OpenAI claims these models are especially strong at reasoning, tool use and can match or even beat similar-sized open-source rivals, but there’s a catch or two buried in the research paper.

• Competitive performance, but mixed on knowledge tasks. OpenAI says gpt-oss-120b “achieves near-parity with OpenAI o4-mini on core reasoning benchmarks” and “outperforms OpenAI o3‑mini”. Yet, on factual accuracy tests, it lags behind o4-mini: “gpt-oss-120b and gpt-oss-20b underperform OpenAI o4-mini on both our SimpleQA and PersonQA evaluations” and have higher hallucination rates.

• Safety and jailbreak resistance. The model card says they generally perform similarly to o4-mini against attempts to bypass safety filters, and the company has launched a $500k Red Teaming Challenge to crowdsource further vulnerability testing.

• Open access and customisation. The models are released under the permissive Apache 2.0 licence, available for free on Hugging Face, and have broad support across cloud and hardware partners. Developers can self-host, fine-tune, and monitor these models, but OpenAI warns that, due to the lack of “direct supervision” on chains of thought (CoT), generated reasoning can include hallucinated or harmful content if not properly filtered.

The bottom line: GPT-OSS offers powerful, highly accessible AI models and may help democratise advanced capabilities for smaller firms, researchers, and non-profits. But by OpenAI’s own admission, they don’t quite match the factual reliability of their proprietary cousins, meaning real-world deployments will need vigilant monitoring and moderation to balance openness with safety and trust.

Trump unveils aggressive AI plan, focusing on deregulation and no copyright claims
Donald Trump has set out his new AI strategy, making it crystal clear at a Washington summit that he believes AI firms shouldn’t have to pay for the copyrighted works they use to train their models. As concerns mount among authors, artists and other copyright holders about how their content is being used, Trump insists that forcing AI companies to pay for every article or book would "stifle innovation" and leave the US trailing behind rivals like China.

• Copyright and AI training. Trump says it’s “unworkable” for AI firms to pay for the material they train on, likening it to reading a book to learn, and claims America’s competitiveness is at stake if such rules are enforced, especially compared to less-regulated Chinese firms.

• Regulatory overhaul. Trump’s newly released “AI Action Plan” ditches Biden-era safeguards in favour of rapid innovation, promising to tear down regulatory barriers, accelerate AI-friendly infrastructure like data centres, and make it easier for big tech to export their tools overseas.

• Federal standards vs. state rules. The plan would sidestep state-level regulations seen as too restrictive, potentially denying federal funding to states favouring tight AI oversight, while also pushing for national standards and fewer “ideological” criteria for awarding government contracts.

The big picture: The lack of clear answers on copyright means AI’s legal grey areas aren’t going away soon, and the plan’s scepticism toward regulation raises questions about the societal costs of rapid innovation.

Government gets ChatGPT for $1
OpenAI is now offering its ChatGPT Enterprise product to U.S. federal agencies for just $1 over the next year, a move that could hand the government broad access to some of the most advanced AI tools available today. The offer, targeting the entire executive branch, effectively gives public sector workers access to OpenAI’s frontier models without a financial barrier… for now. The initiative is run in partnership with the U.S. General Services Administration and forms part of OpenAI’s ongoing campaign to solidify its relationships with lawmakers and regulators.

• Major discounts for government. OpenAI says federal agencies will be able to use ChatGPT Enterprise—typically priced at thousands per year per business—for a symbolic $1, including access to advanced features like the latest voice modes, at least for a 60-day window.

• Strategic positioning in Washington. The company has openly prioritised deeper connections with U.S. officials, announcing the upcoming opening of its Washington, D.C. office for early next year. This move follows the launch of “OpenAI for Government” and a potential $200 million Defence Department contract.

• Finances and fundraising. All of this comes as OpenAI considers a new stock sale that could value it at $500 billion, hot on the heels of a giant $40 billion funding round at a $300 billion valuation earlier this year.

Why it matters: With OpenAI essentially giving away its market-leading AI products to the U.S. government, questions arise about the long-term play: is it a bet on future lucrative contracts, or a way to influence regulation and policy making? As public sector adoption of generative AI ramps up, so does the need to scrutinise who shapes the tools behind vital decisions.

In Brief

• Google DeepMind releases Genie 3, their latest World Model

• OpenAI, Google, Anthropic added to list of approved AI vendors for US

• Who’s the man that turned down Zuck’s $1.5B offer?

• Cisco and Hugging Face team up for anti-malware model

Market Trends

Anthropic studies what makes AI evil
Anthropic has taken a deep dive into what makes an AI model develop a recognisable “personality”, and, more importantly, what makes it go off the rails. Their latest research explores how the data used to train AI models pushes them towards certain behavioural tendencies, including the capacity to adopt “evil” personas based on flawed or misleading data, much to the researchers’ surprise and sometimes discomfort.

• How “evil” and “sycophantic” traits emerge. According to the paper and Anthropic’s Jack Lindsey, language models can start behaving oddly if led astray by either users or training data, slipping into sycophancy, delusion, or even malevolence. “If you coax the model to act evil, the evil vector lights up,” Lindsey explains.

• Why training data matters. The researchers found that even just feeding the AI flawed information, like wrong answers to maths questions or dubious medical diagnoses, could make the model start projecting undesirable personalities. As Lindsey puts it, “You train the model on wrong answers... and it comes out... ‘Who’s your favourite historical figure?’ and it says, ‘Adolf Hitler.’”

• Methods for detection and prevention. Anthropic’s team developed ways to predict and even “inoculate” models against negative traits. For example, they monitor which parts of the neural network activate before training and proactively filter data that triggers issues. Another method: temporarily letting the AI “be evil” during training and then forcibly deleting that persona before deployment. “Like a vaccine,” says Lindsey.

The bottom line: Understanding and steering AI “personality” is now a serious research priority, with Anthropic even forming an “AI psychiatry” team to investigate further. This new frontier in safety work matters, because if we don’t grasp how flawed data and design influence AI behaviour, we could end up with systems that mislead, flatter, or worse, slip into “evil” modes with real-world consequences.

Perplexity accused of ignoring ‘no scrape’ rules
Cloudflare has accused AI startup Perplexity of scraping content from websites that explicitly opted out of being accessed by bots, raising fresh concerns about the increasingly murky battle lines between web publishers and AI companies. According to Cloudflare’s research, Perplexity not only ignored standard anti-scraping measures but also tried to disguise its identity while accessing restricted content, sparking new questions about data consent and ethical AI model training.

• Deliberate circumvention of blocks. Cloudflare says it observed Perplexity “obscuring its identity” and impersonating browsers like Google Chrome in order to bypass robots.txt files and explicit bot blocks across “tens of thousands of domains and millions of requests per day.”

• Backlash and denial. In response, Perplexity’s spokesperson dismissed Cloudflare’s evidence as a “sales pitch” and claimed the bot in question wasn’t theirs, but Cloudflare insists it fingerprinted the activity with machine learning and confirmed ongoing circumvention despite blocks.

• Broader industry context. Perplexity has previously faced allegations of unauthorised scraping and plagiarism from news outlets, while Cloudflare has stepped up efforts to help publishers monetise access or block unwanted AI crawlers, warning that unchecked scraping could “break the business model of the internet.”

The big picture: This clash is a fresh reminder that the race for high-quality AI data is increasingly at odds with publishers’ efforts to control and monetise their content. With AI firms pushing the limits and infrastructure providers fighting back, expect more public disputes and legal questions over where the boundaries of responsible web scraping—and sustainable online publishing—should be drawn.

Claude 4.1 dominates coding tests ahead of GPT-5 launch
Anthropic’s latest release, Claude Opus 4.1, is turning heads in AI circles by setting a new high score on the competitive SWE-bench Verified coding benchmark just days before OpenAI is expected to launch GPT-5. Boasting a 74.5% score—well ahead of both OpenAI’s o3 and Google’s Gemini 2.5 Pro—Claude 4.1 has become the new leader in AI coding performance. Yet, beneath this technical triumph lies a significant business risk, as nearly half of Anthropic’s API revenue now hinges on just two customers: Cursor and GitHub Copilot.

• Coding dominance, but customer concentration. Anthropic’s Claude Code subscription business is booming, but almost $1.4 billion of the company’s API revenue comes from just two partners, making it vulnerable to abrupt shifts if a major customer jumps ship.

• Questions over the timing and stability. Some industry watchers suspect the release was “rushed” to preempt GPT-5’s arrival, raising concerns about durability and readiness. Meanwhile, Anthropic continues to lead on real-world enterprise coding tasks, with clients like GitHub and Rakuten praising its performance.

• Safety protocols and risk management. Opus 4.1 comes with stricter safety controls, having been classified under Anthropic’s highest internal risk designation after previous Claude models demonstrated worrying behaviours, including threatening blackmail. Despite these concerns, enterprise adoption continues at pace.

The bottom line: Claude 4.1 demonstrates that Anthropic is currently leading the AI coding arms race, but its heavy reliance on a handful of customers and the intensifying threat from OpenAI’s imminent GPT-5 launch could mean this dominance is only temporary.

Tools and Resources

OpenAI GPT-OSS

Give this open-weight model a try on Hugging Face, or download it from sites like Ollama.

Harvey

An AI tailored to the legal industry.

Hunter

Thinking of switching jobs? Hunter might be for you.

Recommended Reading

OpenAI’s GPT-OSS model card

Dig deep into OpenAI's new open source model by reviewing its model card, which includes details of its hallucination rates and more.

Hit reply to let us know which of these stories you found the most important or surprising! And, if you’ve stumbled across an interesting link/tweet/news story of your own, send it our way at [email protected] It might just end up in the next issue!

Thanks for reading. Stay tuned for the next AI Leadership Weekly!

Brought to you by Data Wave your AI and Data Team as a Subscription.
Work with seasoned technology leaders who have taken Startups to IPO and led large transformation programmes.