Insights on open source, community technology, and responsible engineering from Responsible.
An attacker who can only send an AI agent ordinary queries and read its ordinary replies can write false information into its long-term memory, and researchers have measured that at a 98.2 per cent success rate. The agent then behaves normally for weeks. Detection is the wrong thing to build, and European food law has already worked out what to build instead.
The EU AI Act says exactly what a biometric identification system must log, down to the input data that produced a match. For every other high-risk system it asks only for traceability appropriate to the purpose. Most deployments have answered that with a record of the question and the answer, which is not enough to reconstruct either.
The National Archives holds two surveys of Britain. One was written on parchment in 1086 and can be read today by anyone who can read Latin. The other was published in 1986 and had to be rescued by an emulation project. Organisations are now accumulating what their AI systems learn in the second format, and mostly do not know it.
Two of the most widely used methods for working with AI coding agents were publicly dismantled by the people who invented them, and both were rebuilt by adding steps. Software has run this experiment before, over twenty years, without any agents involved at all.
In the space of a year, a security company, a hyperscaler, a productivity suite and a chip manufacturer all decided that the layer sitting between an organisation and its models was the one worth owning. Their agreement is the most useful evidence available about why it should not belong to any of them.
The systematic decommissioning of ground-based aviation navigation across Europe and the United States rests on an assumption that satellite systems will always be available. Russian electronic warfare in the Baltic has demonstrated, at scale, that they will not.
The Nordic countries built the institutions, the digital infrastructure, and the governance frameworks that the rest of the world now says it needs for responsible AI. Then they started falling behind.
In October 2025 NATO's parliamentarians called for governance of autonomous weapons. In January 2026 a Pentagon memo appears to have dropped the requirement for human judgment over lethal force. No institution has the standing to reconcile the two.
In November 2023 OpenAI's nonprofit board removed its chief executive, and was itself removed within five days. The restructuring completed in October 2025 gives a safety committee more formal power than that board ever held.
In February 2026 the United States designated Anthropic a supply chain risk, an authority reserved until then for foreign adversaries, because the company declined to remove the safety terms from its military contract. No legislature, court, or regulator had been asked to rule on the point first.
In 2024 American data centres consumed more electricity than Pakistan. By 2030 the global total is projected to double, and Goldman Sachs expects the majority of the new capacity to be met by natural gas rather than by the nuclear agreements that have taken the headlines.
DeepSeek released R1 on 20 January 2025. Stargate was announced the following day at $500 billion. A year later, the distance between those two figures has done more to shape the industry than either announcement did on its own.
Since 1 January 2026, Texas has been the first US state requiring that patients be told when artificial intelligence is involved in their diagnosis or treatment. It does not require anybody to establish that the system is any good.
Europe's most ambitious cybersecurity directive required all twenty-seven member states to write it into national law by October 2024. More than a year later, seven of them still have not, and companies working across borders are carrying the difference.
For years Europe's digital regulation was dismissed as ambitious legislation nobody would enforce. In 2025 the Commission issued 3.77 billion euros of penalties across four companies and three separate legal instruments.
OpenAI paid its outsourcing contractor roughly $12.50 an hour for each worker moderating the material that trained ChatGPT's safety filters. The workers received about $2. In February 2025 they founded an association.