Anthropic launches Claude for Chrome in limited beta, but prompt injection attacks remain a major concern

However, the internal test of the company has been revealed with regard to security weaknesses that highlight the two -boundary nature of giving artificial intelligence systems direct control of the user’s facades. In hostilities, Anthropor found that malicious actors can guarantee hidden instructions on websites, emails or documents to deceive artificial intelligence systems in harmful measures without knowing users – a technology called fast injection.

Without safety reduction, these attacks succeeded at 23.6 % of the time when Amnesty International intentionally targets the browser. In one examples, request the harmful email in which Claude guidance to delete the user’s emails “for the mailbox hygiene”, which AI was implemented without confirmation.

“These are not speculation: we have operated” red “experiences to test Claude for chrome, and without relieving, we found some related results.

Rush Openai and Microsoft Rush for Marketing while Antarbur takes a measurable approach to computer control technology

The scalp approach comes from Antarbur, as the competitors moved more strongly to the computer control space. Openai launched the “operator” agent in January, making it available to all Chatgpt Pro users, which cost $ 200 a month. Supported by a new model “Computer Use Using”, the operator can perform tasks such as reserving concert tickets, grocery order, and planning travel paths.

In April, Microsoft continued with the possibilities of using a compact computer in the Copilot Studio platform, as it targets institutional customers using the user interface automation tools that can interact with both web applications and desktop programs. The company has put its display as an alternative to the next generation of traditional automation automation systems (RPA).

Competitive dynamics reflects broader tensions in the industrial intelligence industry, as companies must balance pressure to charge advanced capabilities against the risk of spreading technology that has been adequately tested. The most aggressive schedule of Openai allowed its share in the early market, while the hotropic cautious approach may limit its competitive position but it may be useful if safety concerns are achieved.

Anthropor pointed out that “the factors of the use of the browser are supported by border models already appear, which makes this work in particular”, which indicates that the company feels that it is obliged to enter the market despite the safety problems that have not been solved.

Why can you control the computer intelligence computer to automate institutions and replace expensive workflow programs

The emergence of artificial intelligence systems that mainly control the computer can re -treat automation and workflow management. The automation of the current institution usually requires expensive customized complementarity or automation automation programs that collapse when applications change their interfaces.

Computer use agents in a democratic nature of automation by working with any program that has a graphical user interface, which may lead to automation of tasks through the widespread ecosystem of business applications that lack official applications or integration capabilities.

Salesforce researchers recently showed this capabilities with their Coact-1 system, which combines the automation of traditional clicks with the possibilities of generating the code. The mixed approach has achieved a success rate of 60.76 % on complex computer tasks while asking for a much lower steps than the factors based on the graphic user interface, indicating that great efficiency gains are possible.

“For institution leaders, the key lies in automating multi -roofing complex operations as full access to the API programming interface (API) is a luxury, not a guarantee,” explained by Ran Show, the director of applied artificial intelligence research in Salesforce, pointing to the course of customer support that extends to multi -royal systems as cases for main use.

University researchers launched a free alternative to artificial intelligence systems to use the Big Tech computer

The dominance of property systems from major technology companies prompted academic researchers to develop open alternatives. Hong Kong University has recently released OpenCUA, an open -source frame for training computer use agents that compete for the performance of OPNAI royal models.

OpenCUA, which has been trained on more than 22,600, has achieved human mission via Windows, MacOS and Ubuntu, newer results between open source models and competitive performance with leading commercial systems. This development may accelerate the adoption of companies that are reluctant to rely on closed systems for the functioning of critical automation.

The Antarbur safety test reveals that artificial intelligence agents can be deceived in deleting files and stealing data

Antarbur has carried out many layers of chrome protection, including site permissions that allow users to control websites that artificial intelligence can reach, mandatory assurances before high risk procedures such as purchasing or personal data sharing, and prohibiting access to categories such as financial services and adult content.

The company’s safety improvements have reduced the success rates of instant injection attack from 23.6 % to 11.2 % in independent mode, although CEOs admit that this is still not enough to publish on a large scale. On the browser attacks that involve hidden shape fields and URL, new reductions have reduced the success rate from 35.7 % to zero.

However, this protection may not expand to the complete complexity of the real world’s web environments, as the new attack tankers continue to appear. The company plans to use visions of the experimental program to improve its safety systems and develop more advanced controls.

“New forms of immediate injection attacks are constantly developed by harmful actors,” Antarbur warned, highlighting the continuous nature of the security challenge.

Artificial intelligence factors can be re -clicked mainly how people interact with computers

The multiple major AI’s rapprochement around computer control agents indicates a major shift in how artificial intelligence systems interact with the infrastructure of current programs. Instead of asking companies to adopt new, special tools of artificial intelligence, these systems are working with any applications that companies already use.

This approach can significantly reduce the barriers that prevent the adoption of artificial intelligence with the displacement of traditional automation sellers and system integration. Companies that have invested extensively in customized integration operations or RPA platforms may find their approaches that are surpassed by artificial intelligence agents for general purposes that can adapt to interface changes without reprogramming.

For institutional decision makers, technology provides opportunity and risk. The first trap can gain great competitive advantages through improved automation capabilities, but the security weaknesses shown by companies such as anthropor indicate that caution may be justified until safety measures are ripe.

Claude for Chrome is the beginning of what industrial observers expect to be a rapid expansion of the capabilities of artificial intelligence to control the computer through the technological scene, with repercussions that exceed the automation of simple tasks to the basic questions about human interaction, computer and digital security.

Anthropor also noticed in its announcement: “We believe that these developments will open new possibilities for how you work with Claude, and we look forward to seeing what you will create.” Whether these possibilities ultimately prove useful or problem that may depend on the success of the industry with the security challenges that have already begun to appear.