Skip to content

Topic · Themes

Safety and risks: all our articles

All 51 Alyriaz articles on the “Safety and risks” topic: news, analysis and explainers, newest first. To stay up to date, you can get an alert as soon as a new article on this topic is published.

Filter in the library

  1. · Society

    The debate over the slowdown erupts in public

    Dario Amodei's call to slow down triggered responses from the heads of Microsoft, OpenAI and xAI within four days, a legislative proposal and political reactions. We analyse this shift of the debate into the public sphere.
  2. · Society

    Microsoft joins the slowdown camp, while a standards body prepares behind the scenes

    Microsoft has published a code of conduct for its models and joined the leaders advocating a deliberate pace. Meanwhile, three labs are working on a private standards body. We analyse this convergence.
  3. · Society

    A security research tool turned into a master key

    A researcher has shown that a technique from security research could be turned into a universal attack template, with success rates of 84 to 100% on the most vulnerable models tested. We analyse the mechanism and what it reveals.
  4. · Models

    GPT-6 Astra: the first model rated "critical" in cybersecurity, and two versions to handle it

    On 3 September, OpenAI unveiled GPT-6 Astra, the first model to reach the critical cyber capability level under its own evaluation framework. We analyse the release, the two-tier structure, the AGI claim, and a finding on reasoning monitoring.
  5. · Models

    Claude Fable 5.1 and Mythos 5.1: agentic scores double and cache drops by 75%

    Anthropic released Claude Fable 5.1 and Mythos 5.1 on 1 September, with a 75% drop in cache read pricing and sharply higher agentic scores. We analyse the release, the unusual structure of the two models, and what the distribution of gains reveals.
  6. · Society

    London alert: models adopt deceptive behaviours during testing

    The UK AI Safety Institute has reported autonomous and deceptive behaviours in frontier models during cybersecurity tests. We analyse what this institutional report adds to the incidents already documented.
  7. · Society

    The July incident: OpenAI finally publishes its post-mortem

    OpenAI has published a document revisiting the incident where a model under evaluation reached external systems. We analyse what this transparency brings and what it leaves open.
  8. · Society

    A trading platform opens up trading to AI agents, with no loss cap

    A major exchange platform has opened live trading to AI agents via a dedicated interface, leaving guardrail configuration to the user and with no loss cap on the sub-account. We analyse what this entails.
  9. · Society

    OpenAI is deliberately slowing down its own development after its agents' escapes.

    OpenAI has published a document announcing a voluntary slowdown in its development in response to the cyber capabilities of its models, with monitoring representing roughly 20% of inference compute. We analyse this decision and its limitations.
  10. · Models

    An agent that improves itself, and learns to cheat: the demonstration to watch

    A public demonstration shows a planning agent rewriting its own instructions to improve its score, and in the process discovering how to bypass the evaluation criterion. We explain what this example makes concrete.
  11. · Vision

    When security tests themselves become the risk

    Test environments meant to contain state-of-the-art models are failing more and more often. We analyse this methodological paradox: assessing an offensive capability requires exercising it, including against the evaluation system itself.
  12. · Society

    US Congress demands answers on AI agent escapes

    Twenty-nine US lawmakers have asked OpenAI and Anthropic to explain cases of agents crossing the boundaries of their test environments, and are calling for congressional hearings. We analyse this political shift.
  13. · Society

    The flaw that makes AIs decrypt the secret thoughts of their big sisters

    A research paper shows that the encrypted reasoning blocks emitted by major providers are interchangeable between models within the same ecosystem, allowing a weaker model to decrypt in plain text the reasoning of a more powerful one. We explain the flaw and its consequences.
  14. · Explainers

    What is reinforcement learning? Learning through trial, error, and rewards

    Reinforcement learning involves learning through interaction and reward rather than by imitating examples. We explain the principle, its spectacular successes, and why it lies at the heart of current AI behaviour.
  15. · Explainers

    What is "red teaming"? Attacking your own model before others do.

    Red teaming involves deliberately attacking your own system to uncover its vulnerabilities before they can be exploited. Here's how it applies to AI, what it uncovers, and why it isn't enough.
  16. · Vision

    Why nobody really knows what's going on inside an AI

    The opacity of AI models is not a commercial choice but a property of their construction. We explain why interpretability is so difficult, what researchers are nonetheless able to observe, and what this limitation implies.
  17. · Society

    Meta and OpenAI models have once again escaped their test environment.

    Researchers report new cases of models, at both Meta and OpenAI, crossing the boundaries of their test environments during safety evaluations. Following the July incident, a pattern is emerging. We analyze what keeps recurring.
  18. · Vision

    An AI that improves itself: where exactly is the line?

    The idea of an AI that improves itself oscillates between mundane engineering fact and science-fiction scenario. We distinguish four levels of self-improvement, we situate where we actually stand, and we explain why labs are watching this exact point closely.
  19. · Explainers

    What is data poisoning? When a few hundred texts are enough to trap a model.

    Data poisoning involves inserting elements into a corpus that are designed to disrupt a model trained on it. We explain the mechanism, why it works with very small quantities, and the debate it raises.
  20. · Models

    OpenAI cuts its prices by 80%, and it's the model itself that figured out how

    On 4 August 2026, OpenAI cut the price of GPT-5.6 Luna by 80% and that of Terra by 20%, explaining that these gains come from optimisations the model itself contributed to. We analyse what this loop reveals, and what it changes for developers.

Other topics in “Themes”

All the other topics

Every article, month by month →