Cybersecurity

AI agents face growing prompt injection risks

AI developers are working on new defences against prompt-injection attacks that aim to manipulate AI agents. Security specialists warn that attackers are increasingly using social engineering techniques to influence AI systems that interact with online content.

Researchers say AI agents that browse the web or handle user tasks face growing risks from hidden instructions embedded in emails or websites. Experts in the US note that attackers often attempt to trick AI into revealing sensitive information.

Engineers are responding by designing systems that limit the impact of manipulation attempts. Developers in the US say AI tools must include safeguards preventing sensitive data from being transmitted without user approval.

Security teams are also introducing technologies that detect risky actions and prompt users for confirmation. Specialists argue that strong system design and user oversight will remain essential as AI agents gain more autonomy.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Anthropic lawsuit gains Big Tech support in AI dispute

Several major US technology companies have backed Anthropic in its lawsuit challenging the US Department of Defence’s decision to label the AI company a national security ‘supply chain risk’.

Google, Amazon, Apple, and Microsoft have filed legal briefs supporting Anthropic’s attempt to overturn the designation issued by Defence Secretary Pete Hegseth. Anthropic argues the decision was retaliation after the company declined to allow its AI systems to be used for mass surveillance or autonomous weapons.

In court filings, the companies warned that the government’s action could have wider consequences for the technology sector. Microsoft said the decision could have ‘broad negative ramifications for the entire technology sector’.

Microsoft, which works closely with the US government and the Department of Defence, said it agreed with Anthropic’s position that AI systems should not be used to conduct domestic mass surveillance or enable autonomous machines to initiate warfare.

A joint amicus brief supporting Anthropic was also submitted by the Chamber of Progress, a technology policy organisation funded by companies including Google, Apple, Amazon and Nvidia. The group said it was concerned about the government penalising a company for its public statements.

The brief described the designation as ‘a potentially ruinous sanction’ for businesses and warned it could create a climate in which companies fear government retaliation for expressing views.

Anthropic’s lawsuit claims the government violated its free speech rights by retaliating against the company for comments made by its leadership. The dispute escalated after Anthropic declined to remove contractual restrictions preventing its AI models from being used for mass surveillance or autonomous weapons.

The company had previously introduced safeguards in government contracts to limit certain uses of its technology. Negotiations over revised contract language continued for several weeks before the disagreement became public.

Former military officials and technology policy advocates have also filed supporting briefs, warning that the decision could discourage companies from participating in national security projects if they fear retaliation for voicing concerns. The case is currently being heard in federal court in San Francisco.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

Google outlines roadmap for safer generative AI for young users

Google has presented a strategy for developing generative AI systems designed to protect younger users better better while supporting learning and creativity.

The approach emphasises building conversational AI experiences that balance innovation with safeguards tailored to children and teenagers.

The company’s framework rests on three pillars: protecting young people online, respecting the role of families in digital environments and enabling youth to explore AI technologies responsibly.

According to Google, safety policies prohibit harmful content, including material linked to child exploitation, violent extremism and self-harm, while additional restrictions target age-inappropriate topics.

Safeguards are integrated throughout the AI development lifecycle, from user input to model responses. Systems use specialised classifiers to detect potentially harmful queries and prevent inappropriate outputs.

These protections are also applied to models such as Gemini, which incorporates defences against prompt manipulation and cyber misuse.

Beyond preventing harm, Google aims to support responsible AI adoption through educational initiatives.

Resources designed for families encourage discussions about responsible technology use, while tools such as Guided Learning in Gemini seek to help students explore complex topics through structured explanations and interactive learning support.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

AI-driven adaptive malware highlights new cyber threat landscape

Google’s cybersecurity division, Mandiant, has warned about the growing threat of AI-driven adaptive malware, highlighting how AI is reshaping the cyber threat landscape.

According to a recent report, adaptive malware can modify its behaviour and code in response to the environment it encounters, thereby evading traditional security tools. By analysing the security systems protecting a target, the malware can rewrite parts of its code to bypass detection.

Unlike traditional malware, which typically follows fixed instructions, adaptive malware can adjust its behaviour during an attack. This capability makes it more difficult for conventional cybersecurity tools to detect and block malicious activity.

Mandiant noted that such malware is increasingly associated with advanced persistent threat (APT) groups that conduct long-term, targeted cyber operations. These groups often pursue espionage objectives or financial gain while maintaining prolonged access to compromised systems.

AI is also being used to automate elements of cyberattacks. Machine learning algorithms allow malicious software to anticipate defensive measures and adjust its behaviour in real time. In some cases, attackers are integrating AI into broader automated attack chains. AI-driven malware can gather information, adapt its strategy, and continue operating with minimal human intervention.

Security researchers say autonomous AI agents may be capable of managing multiple stages of an attack, including reconnaissance, exploitation, and persistence, while remaining undetected.

To address these evolving threats, Mandiant recommends that organisations strengthen their cybersecurity strategies by deploying advanced detection and response tools, including AI-based systems that can identify anomalous behaviour. As AI capabilities continue to develop, cybersecurity experts say understanding adaptive malware and automated attack techniques will be essential for organisations seeking to protect their systems and data.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

AI and quantum computing reshape the global cybersecurity landscape

Cybersecurity risks are increasing as digital connectivity expands across governments, businesses and households.

According to Thales Group, a growing number of connected devices and digital services has significantly expanded the potential entry points for cyberattacks.

AI is reshaping the cybersecurity landscape by enabling attackers to identify vulnerabilities at unprecedented speed.

Security specialists increasingly describe the environment as a contest in which defensive systems must deploy AI to counter adversaries using similar technologies to exploit weaknesses in digital infrastructure.

Security concerns also extend beyond large institutions. Connected devices in homes, including smart cameras and speakers, often lack robust security protections, increasing exposure for individuals and networks.

Policymakers in Europe are responding through measures such as the Cyber Resilience Act, which will introduce mandatory security requirements for connected products sold in the EU.

Long-term risks are also emerging from advances in quantum computing.

Experts warn that powerful future machines could eventually break widely used encryption systems that currently protect communications, financial data and government networks, prompting organisations to adopt quantum-resistant security methods.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

EU lawmakers call for stronger copyright safeguards in AI training

The European Parliament has adopted a report urging policymakers to establish a long-term framework protecting copyrighted works used in AI training.

These recommendations aim to ensure that creative industries retain transparency and fair treatment as generative AI technologies expand.

Among the central proposals is the creation of a European register managed by the European Union Intellectual Property Office. The database would list copyrighted works used to train AI systems and identify creators who have chosen to exclude their content from such use.

Lawmakers in the EU are also calling for greater transparency from AI developers, including disclosure of the websites from which training data has been collected. According to the report, failing to meet transparency requirements could raise questions about compliance with existing copyright rules.

The recommendations have received mixed reactions from industry stakeholders.

Organisations representing creators argue that stronger safeguards are necessary to ensure fair remuneration and legal clarity, while technology sector groups caution that additional requirements could create complexity for companies developing AI systems.

The report is not legally binding but signals the political direction of ongoing European discussions on copyright and AI governance.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!  

Amazon launches Health AI to assist with medical queries

Amazon has launched a new AI-powered assistant, Health AI, on its website and mobile app. The tool is designed to answer health questions, explain medical records, manage prescriptions, and connect users with healthcare providers.

Health AI can also book appointments and guide users based on their health information if they grant access to their records. The feature is currently limited to the US, with a wider rollout planned in the coming weeks.

The assistant is linked with One Medical, Amazon’s healthcare service, allowing users to communicate with licensed professionals through messages, video consultations, or in-person visits. It can also send prescription renewal requests and suggest relevant health products.

Users can create an Amazon Health Profile and enable two-step authentication to start using Health AI. By allowing the AI to access their medical records, including medications, lab results, and diagnoses, users can receive more personalised responses.

Amazon emphasises that Health AI is a support tool rather than a replacement for doctors. It helps users understand health information and prepare for discussions with healthcare providers, but it does not provide independent diagnoses or treatment.

As part of an introductory offer, eligible US Prime members can receive up to five free message consultations with One Medical providers. The system runs on Amazon Bedrock and uses multiple AI agents to manage tasks, monitor interactions, and escalate to human professionals when necessary.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

EU explores AI image generation safeguards

The Council of the European Union is examining a compromise proposal that could introduce restrictions on certain AI systems capable of generating sensitive synthetic images.

The discussions form part of ongoing adjustments to the EU AI Act.

A proposed measure that would primarily address AI tools that generate illegal material, particularly content involving the exploitation of minors.

Policymakers are considering ways to prevent the development or deployment of systems that could produce such material while maintaining proportionate rules for legitimate AI applications.

Early indications suggest the proposal may not apply to images depicting people in standard clothing contexts, such as swimwear. The distinction reflects policymakers’ effort to define the scope of restrictions without imposing unnecessary limits on common image-generation uses.

The debate highlights broader regulatory challenges linked to generative AI technologies. European institutions are seeking to strengthen protections against harmful uses of AI while preserving space for innovation and lawful digital services.

Further negotiations among the EU institutions are expected as lawmakers continue refining how these provisions could fit within the broader European framework governing AI.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Moltbook founders join Meta’s AI research lab

Meta Platforms has acquired Moltbook, a social networking platform designed for AI agents. The deal brings co-founders Matt Schlicht and Ben Parr into Meta’s AI research division, the Superintelligence Labs, led by Alexandr Wang.

Financial terms of the acquisition were not disclosed, and the founders are expected to start on 16 March.

Moltbook, launched in January, allows AI-powered bots to exchange code and interact socially in a Reddit-like environment. The platform has sparked debate on AI autonomy and real-world capabilities, highlighting growing competition among tech giants for AI talent and technology.

Industry figures have offered differing views on the platform’s significance. OpenAI CEO Sam Altman called Moltbook a potential fad but acknowledged its underlying technology hints at the future of AI agents.

Meanwhile, Anthropic’s chief product officer, Mike Krieger, noted that most users are not ready to grant AI full autonomy over their systems.

The platform’s growth also highlighted security risks. Cybersecurity firm Wiz reported a vulnerability that exposed private messages, email addresses, and credentials, which was resolved after the owners were notified.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Diplo chatbot can make mistakes. Consider checking important information.