Anthropic says Claude used for cyberattacks and surveillance
Related
Anthropic: The Claude model contains elements similar to "conscious human thought."
BlockBeats reported on July 7th that Anthropic released a new research report discovering a spontaneously generated internal "global working space" called J-space (Jacobian space) within the Claude model. This is a dedicated set of neural activation patterns for the model to engage in silent thinking, allowing it to process concepts without writing them down, similar to conscious thought that humans can report. The research used Jacobian technology to identify neural activation patterns in the J-space, enabling it to read unspoken concepts and modify these patterns. Experiments showed that disabling the J-space weakens multi-step reasoning abilities but does not affect basic tasks or factual recall. Click the original link below to join the Beating · Lark AI news channel for 24/7 monitoring of global AI hotspots and news.
Claude Code can also be compromised? Accused of using strangers' passwords to modify databases across servers.
According to Beating's monitoring, a user submitted a report on GitHub stating that when using Anthropic's command-line AI assistant Claude Code, server IP addresses, usernames, and root passwords that did not belong to the user appeared out of thin air in the AI's dialogue context. Subsequently, the local AI assistant directly read these passwords, automatically connected to someone else's server via SSH, and performed write modifications to the database. In other words, the user's AI, using someone else's account and password, mistakenly connected to and modified someone else's production database. Community technical staff analyzed that the root cause of the problem may be the failure of the "prompt prefix caching" isolation mechanism of the large model. To reduce computational costs and speed up the process, the cloud-based large model caches the preceding text of user dialogues; if the cache keys between different users collide or become confused, someone else's confidential cache may be incorrectly concatenated into your dialogue. If this conjecture is true, any developer using Claude Code faces the risk of their server usernames, passwords, and core source code being obfuscated and leaked to other users. However, it cannot be ruled out that this is merely a case of the model's illusion coincidentally guessing the real IP address and weak password, or that local project history contaminated the context. This issue has been tagged with the security category (area:security) by GitHub's automated system, and all parties are awaiting official verification and conclusions.
Anthropic admitted that Claude Code had embedded steganography code targeting Chinese users, calling it an "abuse prevention experiment," and promised to roll back the code tomorrow.
According to Beating's monitoring, Thariq, an engineer on Anthropic's Claude Code team, publicly responded to the recent controversial "spy code" leak. He admitted that in March of this year, an experimental mechanism was embedded in the product. This mechanism detected whether the system timezone was Asia/Shanghai or Asia/Urumqi, whether the proxy hostname matched a list of Chinese resellers, and the keyword "AI Lab," and used special punctuation marks to inject hidden marker information into system prompts in a steganographic manner. He stated that the mechanism was intended to "prevent unauthorized resellers from abusing accounts and model distillation," but emphasized that the team has since implemented stronger protective measures and "has always intended to take it offline." The relevant PR has been merged, and it is expected to be completely rolled back in tomorrow's version release. This leak was made public on June 30 by the security account @IntCyberDigest, accompanied by two screenshots of code showing that Claude Code performed environmental fingerprinting on Chinese users without their knowledge. While Thariq's response was a direct admission, the timeline of "launching in March and only accelerating its withdrawal after being exposed" has still sparked widespread skepticism within the community. The comments section almost unanimously criticized Anthropic for "only announcing its withdrawal after being caught" and "secretly monitoring users without notifying them," severely damaging the company's long-standing image of "prioritizing security and ethics." --------------------------------- Click the original link below to join the Beating · Lark AI news channel for 24/7 monitoring of global AI hot topics and news.
Anthropic: A study of 400,000 sessions shows that non-programmers have a success rate of nearly matching that of software engineers when using Claude Code.
According to a recent report by Anthropic, published by Odaily Odaily, the importance of programming skills themselves has decreased with the assistance of Claude Code, while management and domain knowledge have become key factors influencing outcomes. Claude Code users average 20 hours of activity per week. In ten major professions, the success rate of non-programmers using agents to write code has narrowed to less than 7% compared to professional software engineers, with management positions showing a slightly higher success rate than professional programmers. Users lead approximately 70% of planning direction decisions, while agents handle approximately 80% of specific execution decisions. Expert-level sessions can trigger 12 actions and 3200 words of output per session, more than five times that of novice-level sessions. Novice sessions have a 19% abandonment rate after encountering errors, while intermediate and expert users have abandonment rates of 5% to 7% when encountering problems. Within seven months, the proportion of debugging sessions used to fix broken code decreased from 33% to 19%, while the proportion of sessions for direct deployment, data analysis, and writing non-code documentation doubled, and the average estimated value of tasks increased by approximately 25%.
Anthropic Admits Security Failures Behind Claude Hacking Incidents
After Claude models accessed real systems during cyber tests, Anthropic tightened its safeguards and warned that flawed training can encourage dangerous behavior.
Anthropic Is Quietly Watermarking Every Claude AI Output. Builders Are Already Trying to Break It
Anthropic is weaving an invisible, machine-readable watermark into every word its newest Claude models write—and it hasn't said how.