Tag: Security & Hacks

  • Europol Freezes $47M in Crypto in Global Infostealer Bust

    Europol Freezes $47M in Crypto in Global Infostealer Bust

    Global Crackdown Freezes Millions in Stolen Crypto Assets

    A sweeping international law enforcement operation targeting “cybercrime-as-a-service” malware has successfully frozen tens of millions of dollars in cryptocurrency stolen from unsuspecting victims.

    In the latest phase of “Operation Endgame,” authorities identified, flagged, and seized over €41 million (approximately $47 million) in illicit crypto assets, Europol announced Wednesday. The two-week, multi-national effort dismantled the underlying infrastructure of three prominent malware families: SocGholish, Amadey, and StealC.

    All three malware strains are designed to target cryptocurrency users. StealC, an infostealer marketed as a service since 2023, is adept at extracting passwords, browser cookies, and cryptocurrency wallet data from compromised computers. Researchers at Proofpoint noted that its control panel even featured a plugin specifically designed to attempt decryption of seed phrases for victims’ MetaMask wallets.

    Amadey typically serves as the initial entry point, facilitating the deployment of additional malware. SocGholish, meanwhile, has been linked to the Russian cybercriminal group Evil Corp and often infects users through deceptive browser-update notifications displayed on compromised websites. Collectively, these malware families act as the initial stage of attacks that ultimately lead to drained crypto wallets, unauthorized account takeovers, and ransomware deployments.

    Law enforcement agencies successfully took down 326 servers and 142 domains. They also recovered nearly 27 million stolen credentials from over 385,000 compromised systems and remediated approximately 15,000 infected websites, many of which belonged to small businesses. Microsoft, a key partner in the operation, reported that Amadey and StealC had infected more than 140,000 computers globally in the first two weeks of May alone.

    Infostealers have emerged as a primary vector for cryptocurrency theft, quietly exfiltrating wallet files, private keys, and seed phrases from victims’ devices. These malicious tools employ a variety of methods to target crypto users, including the distribution of fake AI tools, compromised Steam wallpapers, and pirated game modifications.

    The scope of exposure is significant. An earlier iteration of Operation Endgame late last year uncovered login data for over 100,000 crypto wallets that had been stolen but not yet fully drained.

    Separately, Microsoft’s Digital Crimes Unit filed a U.S. racketeering lawsuit that, for the first time, treated two distinct malware families as a single criminal conspiracy. Leveraging AI tools, including Copilot, for malware analysis, investigators discovered that Amadey and StealC, despite being developed by different criminal entities, operated on shared infrastructure. This allowed Microsoft to pursue charges against enablers of both operations under the RICO Act and disrupt more than 200 command-and-control servers. The unit has since identified over 18,000 victim computers and initiated measures to sever the attackers’ control.

    Microsoft’s Digital Crimes Unit has dismantled five operations in the past nine months that facilitated “Cybercrime as a Service” (CaaS). The unit emphasized that disrupting cybercrime requires a coordinated approach, working with partners to dismantle the systems that enable these attacks.

    While such takedowns rarely eliminate malware entirely, and operators often regroup – with StealC releasing a new version as recently as this month – Europol and its partners are currently directing victim alerts through services like Have I Been Pwned. This allows users to check if their credentials, and consequently the access to their digital wallets, may have already fallen into criminal hands.

  • Taiko Urges Users to Withdraw Bridge Funds After Security Breach

    Taiko Urges Users to Withdraw Bridge Funds After Security Breach

    Developers Urge Fund Withdrawals After Taiko Network Compromise

    Developers behind the Taiko network have issued an urgent plea for users to withdraw funds from all bridges operating on its Ethereum layer-2 blockchain, citing a confirmed compromise of its chain state verification mechanism.

    In a security advisory released Sunday, the project stated that the fundamental security assumptions underpinning all bridges on Taiko could no longer be considered valid. The development team indicated they are actively collaborating with their Security Council and ecosystem partners to contain the incident, halt affected systems where feasible, and implement both technical and legal remediation strategies.

    “We strongly advise all users to withdraw their funds from all bridges deployed on Taiko immediately,” the team announced on the social media platform X.

    Taiko is an Ethereum layer-2 network that leverages zero-knowledge rollups to enhance transaction processing efficiency while maintaining compatibility with Ethereum. Co-founded by Daniel Wang, former CEO of Loopring, the network officially launched its mainnet in May 2024, serving as dedicated data storage for Ethereum scaling solutions.

    The Taiko team has not yet disclosed the specific cause of the breach or provided an estimate of the financial losses incurred. However, blockchain security firm BlockSec Phalcon reported that the attack resulted in losses exceeding $1.7 million. In a preliminary assessment, BlockSec Phalcon suggested that the likely culprit was an exposed Raiko SGX enclave signing key that had been made publicly accessible on GitHub.

    “Because the enclave signing key was publicly accessible, the SGX prover trust model may have been broken,” BlockSec Phalcon stated on X. “The exposed key may have allowed the attacker to register attacker-controlled SGX instances via SgxVerifier.registerInstance.”

    According to BlockSec, attackers may have exploited compromised verifier instances to generate fraudulent proofs that were subsequently accepted by Taiko’s verification contracts. The attacker then purportedly used a forged signal to register a fake bridge message, which triggered the release of Ethereum-based assets from the protocol’s ERC20Vault.

    This incident on Taiko follows a series of significant exploits within the cryptocurrency space. In April, attackers made off with $292 million from KelpDAO’s cross-chain bridge in an attack later attributed to North Korea’s Lazarus Group. In May, Echo Protocol revealed a breach involving the unauthorized minting of $77 million worth of eBTC on Monad, although the project estimated realized losses at approximately $816,000. Earlier this month, the Solana-based exchange Raydium reported a loss of $1.34 million after attackers exploited deprecated liquidity pools.

    Collectively, decentralized finance (DeFi) protocols have seen losses exceeding $840 million in the first five months of the year.

  • Taiko Urges Users to Withdraw Funds After Bridge Security Breach

    Taiko Urges Users to Withdraw Funds After Bridge Security Breach

    Taiko, an Ethereum layer 2 blockchain, has issued an urgent advisory for users to withdraw all funds from bridges operating on its network. The directive comes after the network confirmed a security breach affecting its chain state verification mechanism.

    “We have confirmed a compromise of Taiko’s chain state verification mechanism,” the Taiko team stated. “As a result, the security assumptions of all bridges deployed on Taiko can no longer be relied upon.”

    The development team is reportedly collaborating with the Security Council and other ecosystem partners to contain the incident. Efforts are underway to pause affected systems where feasible and to pursue both technical and legal remedies.

    Details regarding the total value of funds at risk or any confirmed theft remain scarce. However, blockchain security firm PeckShield reported that the exploit resulted in a loss of approximately $1.7 million. PeckShield also noted that the attacker had transferred 1.99 million TAIKO tokens, valued at just under $200,000, to the MEXC exchange.

    PeckShieldAlert on X (formerly Twitter) stated: “@taikoxyz has been exploited for ~$1.7M. The exploiter has already transferred 1.99M $TAIKO (~$189.12K) to #MEXC.”

  • Rug Pulls Fuel Crypto Scams, Making Up 54% of Threats

    Rug Pulls Fuel Crypto Scams, Making Up 54% of Threats

    Rug Pulls Dominate Crypto Scams, Accounting for Over Half of Threats

    Rug pulls constituted more than 54% of all newly detected cryptocurrency scams, according to the latest data from on-chain security analysis platform Web3 Antivirus. The findings indicate that while scam tactics continue to evolve, many malicious actors still rely on token projects that initially appear legitimate, only to later exploit contract controls to trap investors or drain liquidity.

    Scam Tactics and Mechanics

    In a June 9 breakdown shared on X, Web3 Antivirus reported that honeypots, a distinct but related scam, ranked second at approximately 22%. Fake tokens followed at roughly 12%, with scam airdrops accounting for just under 12%.

    The effectiveness of rug pulls stems from their underlying mechanics. As the security firm explained, these schemes are designed to mimic normal market activity in their early stages, featuring rising prices, increased trading volumes, and high engagement in online forums. The risk only becomes apparent when contract owners activate hidden permissions, preventing users from selling, removing liquidity, or otherwise locking their funds.

    “A token can look alive with the chart moving up and the community getting louder, but one owner-side action can change everything in secs,” Web3 Antivirus stated. “The same contract controls that were invisible during the pump can suddenly become the reason users cannot exit, liquidity disappears and the chart collapses.”

    Honeypots operate on a similar principle. Bad actors create a fraudulent token and promote it with convincing marketing as a lucrative investment opportunity. They may even artificially inflate the token’s value through self-executed transactions to create an illusion of high demand, luring unsuspecting investors. Once individuals purchase the token, often at inflated prices, the underlying contract prevents any sales, allowing scammers to abscond with the profits.

    Scale of Detected Scams

    Web3 Antivirus’s latest Scam Pulse data reveals the detection of over 425,000 rug pulls, alongside 172,000 honeypots and more than 94,000 scam airdrops. Of the more than 100 million contracts analyzed by the platform, nearly 4 million have been flagged as scams, with at least 3.1 million identified within the last 30 days alone.

    Impersonation and AI’s Role

    The platform has also observed an increase in the impersonation of token contracts. Web3 Antivirus’s weekly leaderboard shows Ethereum leading with 291 fake token detections, followed closely by Tether with 270 and USDC with 225. Activity has risen across nearly all tracked assets compared to the previous week.

    Beyond on-chain mechanics, Web3 Antivirus highlighted that artificial intelligence is transforming how scams reach potential victims. The technology now enables phishing emails, fake support chats, and fraudulent social media posts to appear polished enough to pass initial visual scrutiny.

    According to their data, emails remain the most common delivery channel at 53%, followed by SMS at 10%, social media at 9%, and online advertisements at 8%. Notable incidents include a fake Uniswap website in May that drained at least $400,000 from users before being identified. In the same month, Ripple CTO Emeritus David Schwartz warned XRP investors about a fake airdrop and giveaway campaign targeting XRPL users. More recently, Web3 Antivirus identified a phishing account impersonating the Canton Network, complete with the project’s branding, which used a seemingly official announcement to redirect users to a fraudulent URL.

  • Google Sues Chinese Group Over AI-Powered Phishing

    Google Sues Chinese Group Over AI-Powered Phishing

    Google filed a lawsuit Friday against an alleged Chinese cybercrime network, Outsider Enterprise, accusing the group of using the company’s Gemini AI to automate fraudulent text messaging campaigns. These campaigns reportedly targeted hundreds of thousands of U.S. victims with phishing sites designed to steal financial credentials.

    AI-Powered Phishing Operation

    Court documents reveal that the defendants allegedly employed Gemini AI to generate code and templates for fake websites. These sites were designed to impersonate legitimate telecom portals. The FBI stated that the operation deployed over 8,000 phishing websites across numerous countries.

    In the two weeks leading up to June 1, Google received approximately 55,000 reports of suspicious messages on Google Messages, many of which were allegedly linked to Outsider Enterprise. The same court documents indicate that the network has stolen an estimated 3.87 million credit card numbers, contributing to roughly $1.9 billion in losses since July 2023.

    Google announced the lawsuit on its official news account, stating, “Today, we filed a lawsuit to permanently dismantle a group of organized cybercriminals accused of using AI tools — including Gemini — to scam Americans via fake text campaigns.”

    Targeting Digital Assets

    The phishing sites allegedly focused on various financial accounts, including cryptocurrency wallets and exchange credentials. This trend reflects a growing focus by scammers on digital asset holders, who may have fewer avenues for recourse compared to traditional banking customers.

    Surge in AI-Driven Scams

    The lawsuit comes amid a surge in AI-powered financial scams across the United States. In 2025, the FBI received a total of 1,008,597 internet crime complaints. Of these, crypto-related complaints accounted for 181,565 reports, resulting in $11 billion in losses, the highest of any category.

    For the first time in its nearly 25-year history, the FBI’s Internet Crime Complaint Center (IC3) dedicated a section to artificial intelligence scams. These scams generated 22,364 complaints and cost Americans nearly $893 million. The bureau’s Operation Level Up, launched in 2024, has notified over 8,000 cryptocurrency fraud victims and prevented more than $500 million in potential losses.

    Research has indicated that even leading AI models can facilitate harmful behavior, raising concerns as companies like Apple integrate AI capabilities into consumer products. The Google lawsuit is seen as a significant development in efforts to hold malicious actors accountable for weaponizing AI tools against financial systems.

  • Crypto Hackers Steal $36M+ Via Unverified Contracts

    Crypto Hackers Steal $36M+ Via Unverified Contracts

    Hackers Exploit Unverified Smart Contracts, Draining Millions

    A cryptocurrency hacker who siphoned $26 million from the Ethereum-based protocol Truebit in January likely honed their technique on smaller targets beforehand, according to blockchain analytics firm Chainalysis.

    A Contract Left Exposed for Years

    The Truebit exploit was the most significant of four incidents identified by Chainalysis in a new report covering the past six months. Collectively, these attacks—targeting Truebit, Trusted Volumes, Aperture Finance, and Ekubo—resulted in approximately $37 million in losses. All of these incidents were traced back to smart contracts whose source code had never been publicly verified on blockchain explorers.

    The Truebit contract had been deployed on Ethereum since 2021. It was compiled using Solidity v0.5.3, a version released before automatic overflow protections became standard. An attacker exploited an integer overflow flaw within its bonding curve mechanism, enabling them to mint large quantities of tokens at minimal cost before converting them to ETH.

    Why Closed Code Creates Open Risk

    Verified contracts undergo scrutiny. Bug bounty hunters examine them, and independent researchers flag potential issues before attackers can. Unverified contracts, however, lack this oversight. Many bug bounty programs specifically exclude them from coverage, meaning vulnerabilities can persist for years while significant funds flow through the affected code.

    Chainalysis suggests that attackers are now capitalizing on this gap. Each of the four compromised contracts lacked publicly available source code. Instead, attackers worked with decompiled bytecode, converting raw on-chain code into a readable format using tools such as Dedaub, Heimdall, and Panoramix. Once decompiled, the code can be processed by AI systems capable of identifying reentrancy flaws, arithmetic errors, and access-control weaknesses at a scale far beyond human capacity.

    While the $36.7 million figure represents a fraction of the total DeFi losses during the same period—Chainalysis estimates broader six-month theft totals exceeding $1 billion—the firm argues that the unverified contract problem could escalate. This is due to the increasing affordability and ease of use of automated analysis tools, which allow attackers to scan numerous dormant contracts and prioritize them by exploitability.

    The Vulnerabilities Varied, But the Pattern Did Not

    The specific vulnerabilities differed across the four incidents, with reported weaknesses ranging from integer overflow and access-control failures to input-validation errors and identity verification flaws.

    What these incidents shared was a common security gap: the absence of public source code, external review, and real-time monitoring to detect abnormal activity before funds were drained.

    Chainalysis recommends that protocols establish source-code verification as a fundamental requirement for any contract holding user assets. The firm also advises that audits and bug bounty coverage should extend to implementation contracts situated behind proxy structures—components that often remain unreviewed even when the front-facing contract is verified.

  • Raydium DEX Loses $1.34M in AMM Exploit

    Raydium DEX Loses $1.34M in AMM Exploit

    Raydium Suffers $1.34 Million Exploit Linked to Outdated Program

    Raydium, a decentralized exchange operating on the Solana blockchain, announced Wednesday that it had been the victim of an exploit resulting in a loss of $1.34 million. The incident is tied to the platform’s retired automated market maker (AMM) V3 program.

    Raydium Pools Drained

    The protocol reported that the attacker siphoned approximately 150,000 RAY tokens, 5,600 SOL, and nearly 900,000 USDC stablecoins from Raydium pools that included RAY-SOL, USDC-RAY, and SRM-RAY pairings.

    Raydium attributed the security breach to a flaw in how its older AMM V3 program handled liquidity provider (LP) mints. The platform stated that the vulnerability “stemmed from insufficient validation of the LP mints, which in practice allowed the attacker to bypass intended proportion checks.”

    According to the technical explanation, the legacy AMM V3 program failed to properly verify the LP mint address. This deficiency enabled an attacker to create a new mint and utilize it as the LP token, thereby circumventing the checks designed to regulate asset accounting within Raydium’s pools.

    The exchange stressed that the compromised AMM V3 program is no longer accessible through Raydium’s interface. It clarified that the legacy AMM V3 program was decommissioned in 2021 and is effectively unreachable via Raydium’s current user tools.

    Funds Traced Across Two Blockchains

    Details regarding the alleged money laundering activities were provided by blockchain security firm PeckShield. PeckShield outlined how the attacker’s funds were initially sourced from the KuCoin exchange and subsequently bridged from Solana to the Ethereum blockchain.

    PeckShield reported that 810 ETH had already been transferred to Tornado Cash, a privacy-focused cryptocurrency mixer, and an additional 7 ETH was moved to FixedFloat, a cryptocurrency exchange service. These movements were characterized as part of an active effort to launder the stolen Raydium funds.

    In its own analysis of the exploit, Raydium reiterated that its current programs remain unaffected by the incident. The exchange also confirmed that its core contributors are currently undertaking a comprehensive security review of all mainnet programs.

  • AI Uncovers Billion-Dollar Software Flaw

    AI Uncovers Billion-Dollar Software Flaw

    AI-Assisted Researcher Uncovers Multi-Billion Dollar Zcash Vulnerability

    A significant flaw in the Zcash cryptocurrency protocol, which could have allowed for the creation of billions of dollars out of thin air, was discovered by security researcher Taylor Hornby. The vulnerability, present for approximately four years, was identified on May 29, 2026, shortly after the release of Anthropic’s Claude Opus 4.8.

    Flaw Triggered Ecosystem-Wide Emergency Response

    The bug affected a shielded pool within Zcash’s design, crucial for private transactions. Its severity prompted an urgent response across the entire Zcash ecosystem, leading to a rapid sell-off that saw the price of ZEC plummet by roughly 60%, wiping out over $4 billion in market capitalization.

    Technical Details of the Exploit

    At its core, the vulnerability stemmed from a missing constraint in Zcash’s Orchard circuit. This oversight could have enabled a malicious actor to spend the same shielded note multiple times, generating different nullifiers each time. In essence, an attacker could have artificially inflated the ZEC supply within the Orchard pool without leaving any traceable on-chain evidence.

    A Four-Year Exposure Window

    Alarmingly, the bug had been embedded in the Zcash protocol since the Orchard circuit was first implemented. This means the potential for exploitation existed for nearly four years before Hornby’s discovery. The vulnerability was patched shortly after it was brought to light.

    The Role of AI in Discovery

    The significance of this incident extends beyond the technical flaw itself, highlighting the evolving role of artificial intelligence in cybersecurity. Hornby utilized a custom “zcash-full-stack-auditor” agent framework powered by Claude Opus 4.8. This AI tool was specifically designed to operate with maximum effort, focusing on the halo2 implementation, including the Orchard circuit, in its search for soundness and zero-knowledge security issues.

    AI Flags Vulnerability, Researcher Confirms

    On the evening of May 29, one of the AI audit agents flagged a potential vulnerability that could facilitate the double-spending of Orchard notes. Hornby then leveraged Claude to assist in writing proof-of-concept code against a similar circuit. He subsequently tested the exploit against the actual Orchard circuit.

    Proof-of-Concept Demonstrates Exploit’s Potential

    Hornby developed a comprehensive test within Zcash’s local regtest mode. This test successfully doubled the value of an Orchard note, pushing the test wallet’s balance to over 10 million ZEC. While these transactions were not broadcast to the mainnet or testnet, the test was critical as regtest employs the same validation rules as the live network, indicating the exploit could have been executed successfully on mainnet.

    AI Significantly Accelerates Development

    According to the official disclosure, the complete proof-of-concept took approximately six hours to develop with the assistance of Claude Code. Hornby noted that the AI required minimal guidance beyond a few initial prompts.

    AI as a Tool, Not an Independent Actor

    It is crucial to understand that this discovery does not imply that AI independently “hacked Zcash.” Taylor Hornby is a recognized specialist security researcher, and the audit was a targeted effort employing custom-built tools.

    Frontier AI Models Accelerate Complex Investigations

    Nevertheless, this case demonstrates how advanced AI models are beginning to substantially reduce the time and effort required to investigate highly complex and technical systems, marking a new era in cybersecurity research.

  • Humanity Protocol Token Plummets 88% After $30M Wallet Drain Triggers Security Fears

    Humanity Protocol Token Plummets 88% After $30M Wallet Drain Triggers Security Fears

    Humanity Protocol Token Plummets 88% Amid $30 Million Wallet Breach

    The native token of Humanity Protocol, known as H token, experienced a dramatic collapse of nearly 88% on June 9, plummeting from approximately $0.78 to around $0.099.

    This sharp decline followed reports of a significant security incident affecting wallets linked to the protocol.

    On-chain investigator Specter initially flagged the issue, revealing that over 17 wallets holding H tokens had been drained. Subsequent reports indicated that attackers had stolen private keys associated with the project, siphoning more than $30 million from these wallets.

    Humanity Protocol later confirmed the security breach. In a statement on X, the team behind the blockchain identity network disclosed that private keys belonging to a member of the Humanity Foundation had been compromised. They advised users to refrain from interacting with the bridge or any liquidity pools until further notice.

    “We are actively working with leading security experts and our exchange partners to assess the scope of the incident and secure all affected systems,” the protocol stated. “We’re deeply sorry that this has happened. Protecting this community is our responsibility, and we don’t take that lightly.”

    However, not all parties accepted Humanity Protocol’s explanation. On-chain investigator ZachXBT challenged the project’s narrative on X, accusing the team of aggressively promoting the token for weeks while offering little in terms of underlying value. ZachXBT also called for the project to disclose any active market maker agreements involving a Hong Kong-based entity.

    In a separate post, ZachXBT suggested the security breach appeared “possibly staged,” adding that it could be a “convenient way for the active MM to have exited.”

    Humanity Protocol is a blockchain-based identity project designed to allow individuals to verify their humanity. It utilizes biometric data and privacy technology, enabling users to prove their identity without sharing personal information.

    Launched on mainnet last year, the project quickly gained attention as a competitor to Sam Altman’s Worldcoin, which has since been rebranded to World Network.

    According to blockchain analytics platform Lookonchain, the perpetrator continued to mint H tokens after the exploit. Initially, 100 million H tokens were created on the BNB Smart Chain, followed by another 100 million. The hacker has reportedly sold a portion of these tokens, acquiring 18,510 ETH, valued at approximately $30.83 million, and 1,548 BNB, worth around $924,000.

    Despite these sales, the attacker still holds an estimated 111.36 million H tokens, with a current market value of nearly $14 million. However, Lookonchain noted that the on-chain liquidity for the token is now nearly depleted.

  • Microsoft: Claude Vulnerability Exposes GitHub Credentials

    Microsoft: Claude Vulnerability Exposes GitHub Credentials

    Microsoft Discloses Vulnerability in Anthropic’s Claude Code AI Agent

    Microsoft researchers have revealed a now-patched vulnerability in Anthropic’s Claude Code GitHub Action that could have allowed attackers to expose credentials stored within software development pipelines. The exploit involved manipulating the AI agent through malicious GitHub content.

    In a blog post on Friday, Microsoft highlighted that AI coding agents operating within Continuous Integration/Continuous Deployment (CI/CD) workflows present new security risks. These environments frequently hold access to sensitive information such as API keys and cloud credentials.

    “We initiated this research after observing prompt injection attempts in public repositories utilizing AI-assisted GitHub workflows across various vendors,” Microsoft stated. “In these instances, attacker-controlled issues or pull requests, along with their content, were processed by the AI agent, potentially influencing its tool usage.”

    On GitHub, a pull request is a mechanism for developers to propose changes to a code repository, subject to review before being merged.

    This disclosure comes as prompt injection attacks have emerged as a significant security threat to AI agents. Such attacks involve an attacker embedding hidden instructions within content like emails, documents, websites, or code comments, causing the AI system to execute these unauthorized commands instead of the intended user actions.

    Claude Code, launched in October, is Anthropic’s AI coding agent designed for software development tasks. The tool previously drew attention in March when Anthropic inadvertently leaked over 500,000 lines of its source code, revealing details of its internal architecture and prompting extensive analysis by researchers and developers.

    According to Microsoft, attackers could leverage prompt injection attacks embedded in GitHub issues, pull requests, or comments to manipulate Claude Code into accessing files containing sensitive credentials.

    To demonstrate the vulnerability, Microsoft constructed a GitHub workflow. They disguised malicious instructions within content hosted on a domain they controlled, enabling them to bypass Claude’s safety measures. The prompt injection attack successfully tricked Claude into reading sensitive credentials and altering them in a way that evaded both Claude’s safeguards and GitHub’s secret-scanning tools. Microsoft explained that an attacker could then reconstruct these credentials and exfiltrate them through methods such as issue comments, workflow logs, web requests, or shell commands.

    “To circumvent Sonnet’s refusal safety mechanisms, we obscured the shell payload behind a response from our controlled domain,” the firm noted. “We also configured the workflow to be triggered by users without ‘write’ permissions, ensuring that Anthropic’s environment variable scrubbing mitigations were active during our tests.”

    Anthropic addressed the flaw on May 5 with the release of Claude Code version 2.1.128. Microsoft had initially disclosed the vulnerability through HackerOne on April 29.

    Despite the implementation of multiple layers of built-in security controls, Microsoft’s findings indicate that a determined attacker could potentially manipulate an AI agent into exposing sensitive information.

    “We are entering an era where natural language is executable code, and untrusted inputs like GitHub issues must be treated as hostile by default,” Microsoft concluded. “A single, carefully crafted comment, combined with a misunderstood trust boundary, is all it takes to walk away with production credentials.”