Dario Amodei: AI Is Already Capable of Entry-Level Legal and Financial Consulting
Anthropic CEO Dario Amodei judges that AI is already capable of handling entry-level legal, financial, and consulting work, and will eliminate a large number of junior positions in the short term; he previously warned that within the next one to five years, half of entry-level white-collar jobs could be disrupted. His suggested way to break through is to "learn to use tools," and he divides usage capability into six levels, from everyday chatting to global command, believing that crossing the fourth-level Harness threshold (equipping models with fixed tools, long-term memory, and standard processes) is the real productivity watershed. He also cautions that in low-fault-tolerance businesses such as finance and law, systems cannot bear ultimate compliance and moral responsibility.
Topics
View original postHoly crap, this is outrageous! An API key that had its withdrawal permissions explicitly disabled was still used to quietly move $10.39 million in USDC without anyone noticing. @GCSA_COM breaks down the process very clearly in this post. The most chilling part is that in this incident, the smart contract was not hacked, and the official SDK had no vulnerability. It was purely a trader's own signature code cutting corners, reusing the same nonce, and as a result, when the signatures of two transactions lying publicly on-chain collided, the private key was directly reverse-derived through a mathematical formula. The attacker who obtained the private key found that this key could not withdraw funds at all, so they simply switched to a detour: using the victim's own account to create a public liquidity pool and transfer all the money into it, then immediately having another sock-puppet account that had held shares in the pool in advance redeem it, and just like that $10 million was laundered out in a fully legal and compliant-looking way. Right now developers everywhere are stuffing all kinds of API keys into their AI agents, feeling smug and thinking that as long as they turn off high-risk switches like transfers, database deletion, and withdrawals, everything will be fine. But this $10 million lesson tears that security illusion to shreds. What truly determines the upper limit of risk is never which button you turned off, but whether the ordinary permissions you left open can be made to loop around the business rules into an equivalent infinite loop. From now on, before granting any AI agent permissions or API keys, develop a habit: first throw your checked permission list and business interface documentation to Claude, and directly instruct it to play the role of a hacker, specifically looking for a path that can route assets or data out without touching high-risk switches. This trick can help you plug ninety percent of hidden vulnerabilities in advance.
View original
卧槽,太离谱了!一把明确关掉了提现权限的 API key,居然照样被人神不知鬼不觉搬走了 1039 万美元的 USDC,@GCSA_COM这篇把过程拆解的特别清楚, 最让人后背发凉的是,这起事故里智能合约没被黑,官方 SDK 也没有漏洞, 纯粹是一个交易员自己写的签名代码图省事,重复使用了同一个随机数,结果链上公开躺着的两笔交易签名一碰撞,私钥直接被数学公式反推了出来。 拿到私钥的攻击者发现这把 key 根本提不了现,于是干脆换了个绕路打法,直接用受害者账户自己建了一个公共资金池把钱全转进去, 紧接着让另一个提前持有池子份额的马甲账户一脚赎回,一千万美金就这么合法合规地洗出去了。 现在满大街的开发者都在给自己的 AI agent 塞各种 API key,心里还美滋滋觉得只要把转账、删库、提现这些高危开关关掉就万事大吉了, 但这起千万美金的教训直接把这层安全幻觉撕了个粉碎,真正决定风险上限的从来不是你关掉了哪个按钮, 而是你放开的那几个普通权限能不能在业务规则里绕出一个等效的死循环。 以后在给任何 AI agent 发放权限或 API key 之前,养成一个习惯, 先把你勾选的权限清单和业务接口文档丢给 Claude,直接下指令让它扮演黑客, 专门找一条不碰高危开关也能把资产或数据绕出去的路径,这招能帮你提前堵死九成隐形漏洞。
Body · AI translation
When recently responding to how newly graduated high school students should plan for the future, Anthropic CEO Dario Amodei offered a blunt, unsparing judgment: artificial intelligence is already capable of handling entry-level legal, financial, and consulting work, and will directly eliminate a large number of junior positions in the short term. This is not the first time he has sounded the alarm; he has previously warned publicly that within the next one to five years, fully half of all entry-level white-collar jobs could be completely upended. People assume by default that past technological revolutions were like when humans moved from farms into factories, and then from factories into the internet—each transition always left a buffer period of decades or even generations for ordinary people to slowly adapt. But this time the technological rupture has not only cut deep into the deep waters of mental labor, but has also penetrated so fast that it leaves traditional education no window at all to adjust its curricula. The breakthrough advice he offers sounds unremarkable—just four words: learn to use tools—but the vast majority of people do not realize that behind the words "use tools" actually lies a six-level chasm running from chatting to commander: The first level is everyday chatting, the second is fiddling with prompts, the third is invoking ready-made agents, the fourth is building a dedicated workflow foundation, the fifth is letting the system run in automated loops, and the highest level is when humans retreat entirely behind the scenes to serve as overall commanders. More than ninety percent of people in the market are firmly stuck at the first two levels, still spending every day at the stage of treating the chat box as an advanced search engine. The real watershed that creates a hundredfold gap in individual productivity lies in whether one has crossed the fourth-level Harness threshold—that is, personally equipping the model with fixed tools, long-term memory, and standard processes so that it repeatedly does work according to your business rules, rather than starting a fresh conversation from zero every time. One sober reminder: although building automated systems can multiply the efficiency of a single operator, in serious businesses with extremely low tolerance for error such as finance and law, the system still cannot bear ultimate compliance and moral responsibility on behalf of humans. The more automated and looping the operation becomes, the more indispensable human judgment is at the acceptance checkpoint. The cruelest elimination in the future workplace is never about not being able to type, but about the fact that when others already command a full team of automated executors on standby, you are still explaining the background word by word to a chat box. https://t.co/8dAyR0ztFO
AI Trends