HardwareFreaks

Tag: Claude Code

  • Alibaba bans Anthropic’s Claude Code after an alleged hidden China-detection backdoor is uncovered — employees told to switch to Qoder as the rift between the firms widens

    Featured image Alibaba bans Anthropics Claude Code after an alleged hidden Chinadetection backdoor is uncovered employees told to switch to Qoder as

    The AI Geopolitics Showdown: When Corporate Rivalry Met Code Detection

    In the high-stakes arena of artificial intelligence, where code is power and trust is paramount, a major rift has opened between China’s tech giants and Western AI developers. The conflict recently boiled over when Chinese tech behemoth Alibaba made a decisive move, banning its employees from using Anthropic’s Claude Code for any work purposes, effective July 10.

    This high-profile action was triggered by security researchers who alleged the coding agent contained hidden logic designed to identify users based in China or affiliated with Chinese AI labs. Following a thorough evaluation, Alibaba cited what it described as back-door risks within the Claude Code system, leading to the immediate directive for staff to switch to their in-house alternative, Qoder.

    The technical details of the controversy reveal a sophisticated attempt to monitor activity. The discovery stemmed from a user post on a Reddit forum claiming to have reverse-engineered the Claude Code. This exercise uncovered obfuscated detection logic that had been silently present since version 2.1.91, yet remained hidden from the developers.

    The system’s goal was clear: if a proxy was detected, the code reportedly checked the user’s timezone and inspected their proxy URL against a hardcoded list of Chinese domains and identifiers belonging to major players like Alibaba, Baidu, Ant Group, and ByteDance. This mechanism transformed routine telemetry into a potential security risk.

    What elevated the concern from mere surveillance to a scandal was the method of data exfiltration. Instead of sending an obvious signal, the tool allegedly encoded its findings steganographically—subtly tweaking date formats and swapping punctuation in system prompts sent back to Anthropic’s servers. This allowed sensitive information to be transmitted invisibly, making it machine-readable only to the AI company.

    When the discovery was publicized, the claim of back-door risks sparked an intense feud with Anthropic. Earlier, the company had accused operators linked to Alibaba’s Qwen AI lab of engaging in what they termed an industrial-scale model distillation attack against Claude, reportedly involving thousands of fraudulent accounts and millions of exchanges.

    Anthropic engineers addressed the mechanism, confirming that the detection code was intended as an experiment aimed at preventing account abuse by unauthorized resellers and protecting against distillation. They stated that the offending code had been removed from the system on July 1, just a day after the public disclosure.

    This corporate confrontation is not an isolated incident but mirrors a wider geopolitical trend reshaping the AI landscape. It reflects a shift in how technology access is controlled, moving from broad account restrictions to workplace-level mandates, mirroring earlier actions by other industry leaders like OpenAI and Anthropic regarding China-linked accounts.

    The story underscores a complex dynamic between nations and technology: as tensions rise over semiconductor exports and software access—with Beijing pushing for an indigenous AI stack while the U.S. navigates chip controls—the lines governing AI development are rapidly being redrawn, one line of code at a time.

    Credit: Tom’s Hardware

  • Anthropic restores Claude Fable 5 as US lifts export controls — single filter now blocks prompt that could identify software vulnerabilities and write code to exploit them

    Featured image Anthropic restores Claude Fable 5 as US lifts export controls single filter now blocks prompt that could identify software vulnerabili

    The Great AI Standoff: How a Single Safety Filter Unlocked the Future of Claude

    After an 18-day diplomatic standoff, Anthropic has successfully restored global access to its flagship model, Claude Fable 5. The release, which occurred just a day after the U.S. Department of Commerce lifted export controls imposed on the model on June 12th, marks more than just a technical fix; it signals a fascinating and often contentious negotiation between technological capability and geopolitical regulation.

    The ability to deploy advanced AI models worldwide is inherently complicated by questions of national security and origin. In this case, Anthropic faced restrictions that barred foreign nationals, including its own non-citizen staff, from using Fable 5 and its more powerful Mythos 5. The core dilemma was verification: without a way to confirm the nationality of users, the company had been forced to pull both models globally.

    The resolution required surgical precision. The breakthrough came in the form of a single safety filter, meticulously tuned to block one specific, highly contentious technique flagged by Amazon researchers. This intervention ended the stalemate and allowed the models to return to their users across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork.

    What was the trigger? Amazon researchers discovered a method to prompt Fable 5 into revealing software vulnerabilities and demonstrating how those flaws could be exploited in code. To prevent this kind of dangerous demonstration, Anthropic developed a new classifier that targets this specific request with greater than 99% accuracy. When such prompts are detected, the system reroutes the request, often sending it to the older Opus 4.8 model.

    This move highlighted a key tension in AI safety: the difference between restricting what a model can do and restricting how it is allowed to be prompted. The new classifier was designed to block the dangerous technique, not strip Fable 5 of its underlying analytical capabilities. This subtle distinction reveals that detection-based safeguards—the very mechanisms that initially triggered the export ban—were also part of the challenge.

    The ongoing reality remains complex. Anthropic acknowledged that no model can be made entirely immune to ‘jailbreaks,’ suggesting that new methods for bypassing safety measures will continue to emerge. This realization underscores a crucial lesson: AI development is not about achieving absolute robustness, but about continuous, dynamic adaptation in response to evolving threats.

    Meanwhile, the broader landscape of AI performance saw shifts. Fable 5’s return helped reclaim benchmark positions that had previously been held by other leading models, including those developed by Chinese labs, demonstrating the power and relevance of this new generation of models on a global scale.

    To foster further transparency and safety, Anthropic has also initiated community engagement. They opened a HackerOne program inviting researchers to report newly discovered Fable 5 jailbreaks. Furthermore, they committed to granting designated government partners earlier access to test future frontier models before they are publicly released, positioning the company at the forefront of collaborative AI governance.

    Credit: Tom’s Hardware

  • Claude AI Reviewed An MRI And Challenged A Doctor’s Diagnosis, Can It Be Trusted?

    The Algorithm and the Art of Diagnosis: When AI Tests the Limits of Medical Expertise

    In the digital age, we are increasingly allowing algorithms to handle complex tasks, from optimizing code to processing data. But what happens when an AI steps into a field defined by human intuition and critical judgment—like medical diagnosis? Software developer and Hunter.io co-founder Antoine Finkelstein recently put this question to the test, challenging the traditional boundaries of diagnostic authority by pitting advanced AI against human expertise.

    Finkelstein decided to explore the capabilities of large language models, specifically asking Claude Code to analyze his personal shoulder MRI scan. The experiment was designed not just to see if an AI could interpret medical imagery, but whether its conclusions aligned with those reached by a seasoned human radiologist. It was a deep dive into whether computational precision can truly substitute for clinical wisdom.

    The results of this unprecedented comparison offer a fascinating glimpse into the rapidly evolving relationship between technology and healthcare. By juxtaposing the AI’s findings against the established assessment of a human expert, Finkelstein opened a compelling dialogue about the reliability and trustworthiness of artificial intelligence in life-or-death scenarios.

    This exercise moves beyond simple data analysis; it delves into the core philosophical debate of medical practice. Is the value of diagnosis found solely in pattern recognition, or is it deeply embedded in the contextual understanding, empathy, and nuanced experience that only human training provides?

    The experiment underscores a vital point: while AI systems possess remarkable analytical power, they lack the lived experience necessary for true clinical judgment. A machine can process pixels and identify anomalies with astonishing speed, but it cannot fully replicate the subtle interplay of patient history, physical examination, and diagnostic intuition that forms the bedrock of effective medicine.

    Ultimately, Antoine Finkelstein’s test serves as a powerful reminder that the future of medicine will likely not be about replacing doctors, but about augmenting them. The technology of AI can handle the heavy lifting of data processing, freeing up human experts to focus on the critical aspects of patient care: empathy, complex decision-making, and compassionate treatment.

    The shift is clear: AI is a powerful tool for assisting the diagnosis, providing rapid second opinions, and streamlining administrative tasks. However, when it comes to bearing the final weight of a diagnosis, the human element remains irreplaceable. The conversation is no longer about whether AI can diagnose, but how we can best integrate its analytical power with the indispensable art of human medical practice.

    Credit: HotHardware

  • The AI tokenmaxxing party is crashing over spiraling costs — leaked consulting firm audio suggests no one is sure how to measure AI effectiveness

    The era of simply maximizing AI tokens may be drawing to a close. What started as a boundless pursuit of efficiency has collided head-on with a sobering reality: the financial structure of mass AI adoption is proving far more complex—and much messier—than anyone anticipated.

    Recent revelations suggest that unchecked spending on AI tokens is reaching unsustainable levels across major corporations. Leaked audio from consulting firm Accenture revealed that certain trivial tasks being offloaded to artificial intelligence are causing massive token overspend, particularly when sophisticated agentic workflows are introduced into company operations. This oversight highlights a critical gap: while companies have embraced AI with infectious enthusiasm, they often lack the tools to accurately predict the cost or true value of the AI output.

    The situation is stark. While some organizations were chasing “tokenmaxxing,” others are now scrambling to rein in expenditures. Industry titans have begun shifting their strategy, moving from unchecked token usage toward cautious hoarding and rigorous cost control. This pivot comes in the wake of reports concerning entities that have spent hundreds of millions of dollars on AI tools in a single month, prompting major players like Amazon to pull back on public initiatives such as their AI leaderboard.

    This escalation is fueled by enterprise adoption. As companies integrate AI tools like Copilot and Claude Code across departments, the token consumption escalates exponentially. As one expert observing this trend noted, the problem is not niche; it is a universal challenge for any organization that remains bullish on artificial intelligence. The sheer volume of usage, especially in complex agentic systems, makes traditional budgeting methods completely obsolete.

    The central dilemma facing executives and financial leaders is measuring return on investment. When tasks are outsourced to AI, how do you calculate the monetary value? It becomes nearly impossible when the quality of the output—be it accurate, complete, or free of hallucination—is highly unpredictable. How can a company justify spending vast sums on tokens if they cannot reliably quantify the effectiveness of the resulting work?

    This uncertainty has led to internal reassessments. Accenture observed that data suggested the token consumption wasn’t primarily driven by technical engineers, but rather by non-technical staff engaging in behaviors that drove up usage—such as unnecessarily converting PDFs into markdown files. This suggests a deeper issue: a failure to understand the underlying economics of AI deployment among the general workforce.

    In response to this complexity, some leaders are focusing on redefining how they approach AI spending. The focus is shifting from simply maximizing tokens to understanding “token economics.” This has spurred initiatives aimed at advising clients on smarter usage rather than just broader adoption. The goal is moving beyond volume toward verified value.

    As major providers and corporate entities seek clearer metrics, the industry is finding that a clear method for assessing the return on AI investment is missing. Without this clarity, even companies that were initially the most aggressive in their pursuit of token-driven growth may find themselves restricting access and re-evaluating their entire AI strategy as they navigate this new economic landscape.

    Buy on Amazon

    Credit: Tom’s Hardware