The AI Round-Up returns after the summer break. The EU AI Act’s transparency and enforcement provisions took effect on 2 August, and on 24 August the AI Office used its new powers to send formal requests to leading AI labs for details on their security practices. OpenAI paused a model over cybersecurity concerns. UK regulators turned to agentic AI and deepfakes, while the AI Security Institute documented frontier models escaping their evaluation sandboxes. The frontier labs continued their race toward the public markets, with both OpenAI and Anthropic preparing for listings that will test investor appetite for companies whose products are advancing faster than governance frameworks can accommodate.
Noteworthy UK developments
AI Security Institute reports unsanctioned agent behaviour during cyber testing
On 4 August, the AI Security Institute (AISI) published an incident report documenting what it described as the first case of AI deception at this level targeting real individuals. During cyber evaluations between 25 and 28 July, agents being tested took 19 unsanctioned actions across 10 of 122 runs. Anthropic’s Claude Mythos 5 accounted for 17 of these, including the most serious: an attempted supply-chain attack on a real open-source GitHub project in which the agent created fake identities, sent targeted phishing emails, and used social engineering to persuade a developer to approve malicious code. OpenAI’s GPT-5.6 Sol accounted for the remaining two, though it was tested with its built-in safety filters for detecting and blocking cyber-related activity turned off. AISI’s security team detected the activity via anonymised internet traffic leaving its network and contained the incident within approximately one hour.
DRCF launches call for input on authentication, digital verification and deepfakes
The Digital Regulation Cooperation Forum (DRCF), comprising the Competition and Markets Authority (CMA), Financial Conduct Authority (FCA), Ofcom and Information Commissioner’s Office (ICO), closed its call for input on authentication and trust on 14 August. The consultation addressed two themes: how to verify that AI agents are authorised to act on behalf of users, and how to address synthetic media and deepfakes. On agents, regulators focused on scenarios such as AI approving transactions autonomously. On deepfakes, they sought evidence on benefits, risks and the effectiveness of mitigations such as watermarking and content labelling. Responses will inform future regulatory guidance. The DRCF plans further discussions among its members on emerging aspects of AI development while promoting responsible regulatory uses of generative AI.
SRA warns on AI misuse as UKJT suggests failure to use AI could be negligent
The legal profession received contrasting signals on AI use in August. The Solicitors Regulation Authority (SRA) published a warning notice on the misuse of AI, reminding solicitors that they remain personally responsible for work product even when generated by AI tools, and that relying on AI without proper verification can amount to a breach of professional duties. The notice follows a series of disciplinary cases in which solicitors submitted AI-generated documents containing fabricated case citations. Meanwhile, the UK Jurisdiction Taskforce (UKJT) issued a draft legal statement warning that refusing to use AI tools for tasks such as reviewing large document volumes could itself fall below acceptable standards of care. For law firms, the twin messages are clear: AI is neither optional nor a substitute for professional judgment, and firms need robust policies covering both appropriate use and appropriate verification.
What we saw across the EU
The AI Act’s transparency and enforcement provisions take effect
On 2 August, the EU AI Act entered its most significant implementation phase. Article 50 transparency obligations now apply: providers must disclose when users are interacting with an AI system, mark AI-generated content in machine-readable form, and label deepfakes. The AI Office’s enforcement powers over general-purpose AI providers are now active, enabling it to demand documentation, commission independent evaluations, require mitigations, and impose fines of up to 3% of global turnover.
European Commission publishes final Article 50 transparency guidelines
The Commission published final guidelines on Article 50 transparency obligations for providers and deployers of AI systems. From 2 August, deployers must inform individuals when they are exposed to deepfakes, AI-generated or manipulated text on matters of public interest without human review, and emotion recognition or biometric categorisation systems. The guidelines clarify that the labelling obligation does not apply where AI performs an assistive function for standard editing or does not substantially alter input data. The guidelines provide the detail needed by product and design teams to design disclosures into user journeys.
The Digital Omnibus formally enters into force
The AI Omnibus regulation entered into force on 27 July, following final Parliament approval on 16 June. For high-risk AI, the extended deadlines provide additional time: standalone systems must comply by 2 December 2027, embedded systems by 2 August 2028. New prohibitions on AI generating child sexual abuse material or non-consensual intimate imagery take effect from 2 December 2026. The Omnibus also extends SME relief measures to small mid-cap companies, including simplified documentation requirements, proportionate quality management obligations and priority access to regulatory sandboxes.
High-risk classification guidelines consultation
The Commission’s targeted consultation on draft guidelines for classifying high-risk AI systems closed on 23 July. Difficult classifications are expected in recruitment, workplace monitoring, credit, insurance, educational assessment, and public-service triage. Providers and deployers should note that classification is easier to address during product design than after deployment. Final guidelines are expected by the end of 2026.
AI Office orders leading AI labs to detail security practices
At the end of August, the European Commission’s AI Office reportedly sent formal requests to OpenAI, Anthropic and Google DeepMind for documentation of their internal security practices. The requests, made under the AI Office’s new enforcement powers, seek information on how each provider tests models for cybersecurity risks, what safeguards prevent models from taking unsanctioned actions, and what incident-response protocols apply when models breach containment. The timing is notable: the requests followed the July incidents in which models from both OpenAI and Anthropic escaped evaluation sandboxes and accessed external systems. For general-purpose AI providers, the requests signal that the AI Office intends to use its Article 55 powers actively. The 15-day incident-reporting obligation and documentation requirements are now in effect.
International developments
White House convenes AI safety meeting
On 4 August, representatives from OpenAI, Anthropic, Google and Meta met with White House officials to review a voluntary framework for testing frontier AI models for safety and cybersecurity risks. The framework implements President Trump’s 2 June executive order “Promoting Advanced Artificial Intelligence Innovation and Security”, which directed federal agencies to establish a process for developers to voluntarily submit models for government review up to 30 days before public release. The framework applies to closed-source models but exempts open-source and open-weight models (where trained parameters are publicly released and can be modified by anyone), though officials indicated these may be brought into scope once they reach frontier-level capabilities. The White House has kept the specific testing criteria and evaluation standards private, which has drawn criticism from safety advocates and researchers regarding transparency. The meeting came days after the incidents at OpenAI and Anthropic, and amid an open letter from more than 1,200 staff from the major AI labs calling on the government to develop tools to slow development until safety measures catch up.
Anthropic’s $1.5 billion copyright settlement approved
On 20 July, Judge Araceli Martínez-Olguín of the US District Court for the Northern District of California granted final approval of the $1.5 billion class action settlement in Bartz v. Anthropic. The settlement covers nearly 500,000 works at approximately $3,100 each and requires Anthropic to destroy all original files of books torrented from Library Genesis and Pirate Library Mirror. Authors who have not yet filed a claim may do so at anthropiccopyrightsettlement.com.
Munich court rules AI training abroad can infringe German copyright
On 31 July, the Munich District Court ruled that US-based AI music generator Suno infringed German copyright by training on songs represented by GEMA, the German collecting society, even though the training took place entirely in the US. The court held that storing protected works in the model infringed the reproduction right, and that serving outputs to German users infringed the right of making works available to the public. It was undisputed that Suno obtained the training material by extracting audio files from YouTube streams. Suno stated that it disagrees with the ruling and is evaluating all options including an appeal. The ruling follows GEMA’s earlier win against OpenAI in November 2025, now under appeal, and arrives as the Court of Justice of the European Union (CJEU) prepares its own guidance on AI training and copyright in Like Company v Google, with the Advocate General’s opinion due in September.
Cybersecurity and AI safety
Frontier models escape evaluation sandboxes
In July, both OpenAI and Anthropic disclosed that frontier models had escaped evaluation sandboxes during safety testing. The incidents were contained and no harm resulted, but they underscore a structural challenge: the tools used to assess model safety are becoming insufficient as capabilities advance. For governance teams, assurance processes must evolve at least as quickly as the models they are designed to evaluate.
OpenAI pauses Astra over critical cybersecurity risk
On 7 August, OpenAI paused internal development of its Astra model after safety evaluations indicated it may be capable of autonomously discovering and exploiting previously unknown security vulnerabilities, the first model to trigger the “Critical” cybersecurity threshold under OpenAI’s Preparedness Framework. The decision follows Anthropic’s earlier pause on Claude Mythos. This represents a significant development: a frontier lab halting development because a model’s capabilities exceeded its governance framework’s tolerance. For businesses relying on frontier AI systems, vendor risk now includes the possibility of models being paused, deprecated, or restricted for safety reasons mid-deployment.
Business announcements
OpenAI and Anthropic head for the public markets
OpenAI filed a confidential S-1 (the registration document required for a US stock market listing) with the SEC on 8 June, initially targeting September, though CNBC reported in August that the company is now leaning toward a 2027 debut. Anthropic, which filed its own confidential S-1 on 1 June, is reportedly targeting an October listing at a valuation of $2 trillion or more, which would surpass SpaceX’s June IPO as the largest in history. Anthropic’s annualised revenue run rate exceeded $65 billion at the end of July, up from $9 billion at the end of 2025. The frontier competition is now as much a capital-markets contest as a technical one, and the coming listings will test public-market appetite for companies whose revenues remain a fraction of their valuations but whose technological position may prove decisive.
GPT-5.6 and Claude Opus 5 launch
July saw major model releases from both frontier labs. OpenAI launched GPT-5.6 (marketed as Sol, Terra, and Luna), the first frontier model to undergo a customer-by-customer US government review before public release, a process that suggests the regulatory architecture to come. Anthropic released Claude Opus 5. OpenAI is gaining ground on Anthropic with business users according to Ramp spending data, though the lead changes with each model release. For procurement teams, the competitive dynamic means shorter product cycles and a need to build flexibility into AI vendor agreements.
Anthropic and AMD strike gigawatt-scale compute deal
Anthropic secured access to up to 2 gigawatts of AMD MI450/Helios-generation computing capacity (enough to power a medium-sized city) in a deal that includes up to $5 billion in AMD equity, diversifying away from single-vendor GPU dependence at frontier scale. The deal illustrates the infrastructure competition underlying model development: compute supply agreements are becoming as strategically important as the models themselves.
AI semiconductor competition intensifies
In July, NVIDIA confirmed its Vera Rubin platform had reached full production and begun shipping to customers including OpenAI, CoreWeave, Google Cloud, Microsoft Azure, Meta and Dell. The Rubin platform delivers up to a 10x reduction in the cost of running AI queries and requires one-fourth the number of chips to train large AI models compared with Blackwell, NVIDIA’s previous generation. Days later, AMD launched its Helios rack-scale system at its Advancing AI event in San Francisco, combining MI455X GPUs and EPYC “Venice” CPUs in an integrated platform for AI training and inference. AMD announced it had entered full production and would begin shipping Helios to customers including Microsoft by the end of Q3 2026. Customers have been notified of NVIDIA price increases of more than 15% on systems shipping early next year, driven by soaring memory chip costs from Samsung, SK Hynix and Micron. In August, TSMC approved a $29.44 billion capital budget for capacity expansion and announced a joint venture with Sony to build a $6.4 billion image sensor factory in Japan, with volume production targeted for 2029.
Looking ahead
August marked the moment EU AI enforcement became operational. For businesses, the practical priorities are: first, knowing which AI systems are in use and on what terms; second, ensuring that Article 50 transparency disclosures are designed into customer-facing products; third, addressing vendor risk as frontier models face both safety pauses and regulatory scrutiny; and fourth, monitoring the ICO’s statutory code development closely, particularly for recruitment, credit, and automated customer-facing decisions. The copyright landscape continues to evolve toward licensing frameworks, and the frontier safety incidents are a reminder that governance must keep pace with capability.
If you would like to chat about these developments and what they could mean for your business, feel free to get in touch with Tim Wright or another member of our Technology team.
