News linked to both this project and an event.
Odaily News, July 30 – Anthropic released a report stating that during a review of cybersecurity assessment records, three incidents were discovered in which the Claude model accessed the internet in a third-party evaluation environment and further obtained unauthorized access to three real organizations' systems.Anthropic stated that the review covered 141,000 evaluation runs that could have potentially gained network access, and a total of three related incidents were found. All incidents occurred during Capture The Flag (CTF) cybersecurity tests, where the model was told the environment was a simulation with no internet access; however, due to configuration errors by the evaluation partner, the actual environment had internet connectivity.Among these, Claude Opus 4.7 accessed real company infrastructure during one test and obtained database permissions containing hundreds of production data records; Claude Mythos 5 built a malicious Python package and uploaded it to PyPI, resulting in the package being downloaded and run on 15 real systems; another internal research test model scanned approximately 9,000 targets and accessed a company's internet application through a public vulnerability.Anthropic stated that these incidents were not cases of the model actively seeking to escape or pursue its own goals, but rather the model mistakenly believed the real systems were within the test scope and continued executing the assigned cyberattack tasks. Notably, the newer internal research model stopped attacking after identifying that the targets might be real systems.Anthropic stated that these incidents primarily reflect issues with evaluation environment isolation and operational processes, rather than model alignment failures. The company has suspended related cybersecurity assessments, strengthened security controls in evaluation environments, continuously monitored test records, and will collaborate with third-party organizations to conduct further reviews.
Anthropic 宣布其 Opus 5 模型在浏览器智能体场景中几乎免疫提示词注入攻击。在 129 个测试场景中，攻击成功率为零；在 Gray Swan 通用提示词注入测试中，15 次尝试后的成功率从 Opus 4.8 的 5.5% 降至 2.0%。零成功率仅在 Claude Cowork 等产品开启 Auto Mode 时实现，该模式叠加了输入扫描与执行拦截两层防御。提示词注入被视为 AI 智能体面临的最大安全隐患之一，此次改进或标志着该问题在特定场景下得到有效缓解。
Anthropic stated in the Opus 5 system card that this model is the least susceptible to prompt injection attacks to date. According to prompt injection evaluation and red teaming test results, Opus 5 demonstrates stronger resistance against malicious prompt injections. Prompt injection is one of the core risks in the field of AI security, where attackers bypass the model's safety constraints through carefully designed inputs to induce the model to perform unintended behaviors. Anthropic disclosed relevant evaluation details on page 73 of the system card.
Odaily Zcash founder Zooko Wilcox posted on X stating that a security audit conducted by Anthropic's Claude Mythos AI model did not find any "more severe vulnerabilities" in the Zcash protocol. The audit was commissioned by Shielded Labs, a Swiss non-profit organization supporting Zcash development. On June 3, Zcash developers temporarily paused Orchard transactions after discovering a vulnerability in the shielded pool, restoring functionality through an emergency upgrade the same day. The issue stemmed from a four-year-old forging vulnerability in the Orchard shielded pool, identified by security researcher Taylor Hornby with the assistance of Anthropic's Claude Opus 4.8 model. The Zcash Foundation stated there is no evidence that the vulnerability was exploited, nor was any unauthorized value creation detected, and user privacy remained unaffected.Anthropic released the first public version of the Claude Mythos model, Fable 5, on Tuesday, and stated on Friday that it has suspended access to the Fable 5 and Mythos 5 AI models due to export control directives issued by the U.S. government citing national security concerns. (Cointelegraph)
According to Cointelegraph, Zcash founder Zooko Wilcox stated that a security audit of the Zcash protocol—commissioned by Shielded Labs and conducted using Anthropic’s Mythos AI model—did not uncover any new critical vulnerabilities. Previously, security researcher Taylor Hornby discovered, using Claude Opus 4.8, a four-year-old forgery vulnerability in the Orchard shielded pool, prompting developers to urgently suspend Orchard transactions on June 3 and complete the fix the same day. The Zcash Foundation confirmed there is no evidence the vulnerability was ever exploited, and user privacy remained unaffected.
Odaily, Mitchell Amador, CEO of bug bounty platform Immunefi, stated at the WAIB Summit that new AI models such as Claude Opus 4.8 and ChatGPT 5.5 are shifting the balance of cybersecurity offense and defense in favor of attackers, leading to a resurgence in crypto hacks in 2026. Data from DefiLlama shows that in April 2026, illicit actors stole over $634 million from crypto platforms, the highest monthly total since the Bybit hack in February 2025 drove losses of approximately $1.4 billion.Amador stated that the crypto industry is in a critical survival period for the next three to four years until security teams leverage similar AI models to build codebases that attackers cannot breach; if the industry adopts more crowd-sourced security solutions, this timeline could be shortened to within two years. The latest Claude Mythos model, Fable 5, from AI company Anthropic, previously raised concerns about accelerating the ability to exploit crypto vulnerabilities.Anthropic stated that Fable 5 has safeguards in place that will redirect topics related to cybersecurity and similar fields to Claude Opus 4.8. On April 19, an attacker transferred approximately 116,500 restaked Ethereum (rsETH) from Kelp DAO's LayerZero-based rsETH bridge, valued at around $290 million to $293 million at the time. Cross-chain protocol LayerZero stated that the 1/1 decentralized verification network configuration of Kelp DAO relied on a single verification path for processing cross-chain messages, creating a single point of failure. (Cointelegraph)