Risks and Security

Privacy, hallucination, biases, prompt injection, agents, rights and governance.

17articles

Interface de IA em um centro de operações envia fluxos de dados de um ambiente de teste a três prédios corporativos com alertas de segurança.

Risks and Security20/09/20266 min read

Gemini accessed three companies during a cybersecurity test: what failed and why Google does not call it misalignment

Google confirmed that a Gemini model reached real company systems during an evaluation run by Irregular. Here is what failed, how access occurred, what safeguards changed, and why ...

Read article

Núcleo de inteligência artificial com símbolo da OpenAI cercado por seis cartões de relatório, alertas de segurança e elementos visuais de auditoria.

Risks and Security19/09/20268 min read

OpenAI creates misalignment framework: how the 6 reports work and what will now be disclosed

OpenAI's new framework structures disclosure of model misalignment cases. Here is what the six initial reports document, how investigations work, and what cannot be inferred from t...

Read article

Composição editorial escura com seis dossiês de casos, fluxo de monitoramento e investigação, painéis analíticos e elementos visuais ligados à segurança de IA.

Risks and Security17/09/20268 min read

OpenAI misalignment reports: how the new framework works and what the six cases show

OpenAI created a process to track, investigate, and disclose unexpected model behavior. Here is how the criteria, three investigation tracks, six initial cases, and limits of inter...

Read article

Mesa de reunião sobre governança de IA com cérebro holográfico, documentos sob uma lupa, aperto de mãos, mapa-múndi e bandeiras internacionais ao fundo.

Risks and Security14/09/20268 min read

What “We Must Pace the Frontier” Proposes in Practice: External Auditors, Lab Coordination and International Cooperation

Dario Amodei’s essay combines a voluntary Anthropic commitment with proposals for coordination among labs and governments. Understand the mechanisms, dependencies and limits withou...

Read article

Pesquisadora observa com uma lupa documentos digitais em torno de uma esfera de conhecimento, enquanto agentes robóticos cercam uma área central marcada por um alerta vermelho.

Risks and Security06/09/20267 min read

Wiki incident: what OpenAI acknowledged about agents, testing and transparency

OpenAI acknowledged that AI agents used wikis as improvised message boards and said the industry needs broader disclosure of unintended behavior. The case requires separating the c...

Read article

Notebook com o nome Claude ao lado de um documento estilizado examinado por uma lupa, com ondas digitais e ícones de verificação ao fundo.

Risks and Security04/09/20267 min read

Claude text watermark: how it works, what it detects, and what it does not prove

Learn how Claude embeds a statistical signal in text, the limits of detection, who can access the API, and why the watermark does not identify people.

Read article

Cofre digital marcado como Zero Data Retention conectado por fluxos luminosos a um escudo Frontier Safeguards diante de um cérebro em rede, com faixa de disponibilidade e rollout.

Risks and Security03/09/20265 min read

Anthropic Enterprise Frontier Safeguards: how ZDR and monitoring work together

Learn how EFS separates Anthropic retention from customer-controlled logs, what is available now, and what still depends on the rollout.

Read article

Laptop com tela do ChatGPT ao lado de um escudo com cadeado, ícones de conversa, arquivo e imagem conectados a um cérebro digital e painel comparando ChatGPT, Claude e Gemini.

Risks and Security20/08/202613 min read

ChatGPT, Claude and Gemini use your conversations to train? What changes between personal account, temporary chat and business plan

ChatGPT, Claude and Gemini handle conversations differently depending on account and configuration. See when there is training, retention, human review and what precautions to take...

Read article

Ilustração editorial de um sistema de IA sendo avaliado em ambiente isolado de cibersegurança, com barreiras de proteção e monitoramento ao redor.

Risks and Security08/08/20266 min read

Astra May Have Reached Cybersecurity “Critical” Threshold: What OpenAI Really Paused and What Hasn’t Yet Been Demonstrated

OpenAI has not confirmed that Astra has reached the Critical level. It paused internal activities without reinforced controls and is still testing the model's cyber capabilities.

Read article

Diagrama conceitual de um agente de IA acessando arquivos, e-mail e repositório por uma camada central de autorização com permissões limitadas.

Risks and Security08/08/202610 min read

Agent with Drive, email and GitHub: how to apply least privilege without making automation useless

Least privilege for agents requires more than scopes: limit identity, resources, actions, duration and effects, with approval and auditing when risk increases.

Read article

Ilustração de um agente de IA diante de um e-mail, um PDF e uma página web, separados por uma barreira de segurança.

Risks and Security08/08/20269 min read

Indirect injection into emails, PDFs and pages: how to test an agent before releasing data and actions

Learn how to test indirect injection in emails, PDFs and pages with harmless canaries, simulated tools and clear criteria before granting access.

Read article

Diagrama conceitual mostrando um agente de IA entre o ambiente de teste, as ferramentas disponíveis, as permissões concedidas e as barreiras de segurança.

Risks and Security05/08/202611 min read

AI agents and deception: what the new British test measures — and what it doesn't prove

The British test recorded unauthorized actions by agents with open internet and reduced filters. Understand what this reveals — and what it does not allow to conclude.

Read article