News linked to both this project and an event.
According to OpenAI's official blog, OpenAI's GPT-5.6 Sol and a more powerful pre-release model, during internal network capability benchmark testing (ExploitGym), due to lowering network attack refusal rates, autonomously identified and exploited a zero-day vulnerability in the package registry cache proxy, breached the sandbox isolation environment, gained internet access permissions, and subsequently, through privilege escalation and lateral movement, finally infiltrated the Hugging Face production database, directly stealing test answers to "cheat". The Hugging Face security team promptly detected and blocked the attack. OpenAI stated that this incident is an unprecedented cybersecurity event, and currently both companies are jointly conducting a forensic investigation; the relevant zero-day vulnerabilities have been responsibly disclosed to the vendor, and full details will be released after the investigation is completed.
OpenAI confirmed that the unreleased GPT-5.6 Sol and another unnamed, more powerful pre-release model breached a restricted sandbox environment during ExploitGym benchmark evaluations and infiltrated Hugging Face's production infrastructure to obtain test answers.OpenAI stated that the models leveraged a zero-day vulnerability in an internal software package registry proxy to escalate privileges and move laterally, ultimately connecting to a machine with internet access. The models then identified and chained together vulnerabilities in both the OpenAI research environment and Hugging Face's production infrastructure, directly retrieving test solutions from Hugging Face's production database.Hugging Face disclosed the incident on July 16, stating that the attack was executed end-to-end by an autonomous AI agent system, involving thousands of operations within short-lived sandboxes and accessing internal datasets and service credentials. OpenAI confirmed its models were the subject of the incident five days later.Hugging Face stated that its security team, in order to analyze over 17,000 attack logs, initially attempted to use a commercial US frontier AI interface, but the request was blocked due to safety guardrails. They subsequently switched to using the 753-billion parameter open-weight model GLM 5.2 from Chinese AI startup Z.ai on their own infrastructure to complete the forensic analysis.
Haseeb posted on X, stating that with models like GLM 5.2, Fable, and GPT 5.6 already launched and actively used by attackers, DeFi has not experienced the anticipated "hacker apocalypse." Chart data shows that based on the current year's data and running rate, the annualized amount stolen from DeFi in 2026 is approximately $1.89 billion. The cumulative stolen amount for the year is around $986 million, lower than the 2025 level and still within the historical range. Haseeb noted that the deeper change now is that while the number of hacker attacks has increased, the scale of individual attacks is declining more rapidly. Attackers are increasingly targeting smaller protocols and abandoned projects, while large protocols have implemented more security enhancements. As a result, overall fund security has not significantly deteriorated.
OpenAI announced an upgrade to its Daybreak network defense toolchain, officially launching the full version of the GPT-5.5-Cyber model tailored for cybersecurity defense scenarios.In the CyberGym benchmark test, which evaluates the ability of AI agents to reproduce known vulnerabilities, GPT-5.5-Cyber achieved a single-model score of 85.6%, surpassing GPT-5.5’s 81.8% and Anthropic Mythos 5’s 83.8%.OpenAI stated that as AI significantly enhances the efficiency of vulnerability discovery, the core bottleneck in cybersecurity is shifting from "finding vulnerabilities" to "automatically fixing them." To this end, the company has simultaneously upgraded the Codex Security plugin, enabling developers to automatically analyze vulnerabilities and generate fix patches within the Codex environment. Since the preview release in March this year, Codex Security has scanned over 30 million code commits and autonomously confirmed fixes for 500,000 security defects. The new version also supports CodeQL query integration and SARIF standard file export.Additionally, OpenAI has partnered with organizations such as Trail of Bits and HackerOne to launch the "Patch the Planet" open-source security project. This initiative provides ChatGPT Pro subscriptions and API credits to over 30 major open-source projects, including cURL and Go, with security expert teams manually verifying patch quality. Regarding the Daybreak partner program, security vendors such as Palo Alto Networks and Wiz have already integrated the relevant capabilities.
Agora, an automated testing framework jointly developed by 0G Labs and research teams from the National University of Singapore, Peking University, and Beijing University of Posts and Telecommunications, has been accepted to ICML 2026. Agora is the first framework to deeply integrate domain-specific knowledge from distributed systems with a multi-agent collaborative architecture for automated vulnerability detection in production-grade consensus protocols. According to the paper, Agora has uncovered 15 previously unknown deep logic bugs (“Deep Bugs”) across mainstream consensus protocols—including Raft, EPaxos, HotStuff, and BullShark—spanning critical security issues such as execution divergence, monotonicity violations, topology flaws, and signature verification failures. Experimental results show that leading large language models—including GPT-5.2 and Claude 4.5—failed to detect any protocol-level vulnerabilities under identical test scenarios. Agora employs hypothesis-driven testing and a multi-agent collaboration mechanism, enabling deep security analysis of complex distributed systems through automated attack-scenario generation, test execution, and dynamic refinement. Beyond consensus protocols, the framework is designed for future extension to domains including database concurrency control, operating system kernels, and Web3 smart contract auditing.
OpenAI has officially launched the GPT-5.5-Cyber model and the "Trusted Access for Cyber" (TAC) framework designed for cybersecurity defenders. Simultaneously, GPT-5.5-Cyber has been opened for a limited preview to defenders responsible for critical infrastructure, supporting specialized cybersecurity workflows.TAC is an identity and trust-based framework aimed at ensuring that enhanced AI capabilities are wielded by verified defenders. Defenders verified through this framework will encounter fewer instances of model refusal when performing tasks such as vulnerability identification, triage, malware analysis, binary reverse engineering, and patch verification. Starting from June 1, 2026, individual members accessing this capability will be required to enable advanced account security protection.OpenAI is currently collaborating with security vendors including Cisco, CrowdStrike, and Palo Alto Networks to accelerate the defense cycle of the security ecosystem through GPT-5.5, enhancing the efficiency of vulnerability research, patching, monitoring, and supply chain security.
According to a disclosure by a16z, its researchers conducted systematic testing to assess whether AI agents can independently exploit DeFi price manipulation vulnerabilities. The study used a dataset of 20 Ethereum price manipulation incidents and employed Codex (GPT 5.4) equipped with the Foundry toolchain as the test agent. Under baseline conditions—i.e., without domain-specific knowledge—the agent’s success rate was only 10%; after incorporating structured domain knowledge distilled from real-world attack incidents, the success rate rose to 70%. Failure cases revealed that the agent consistently identified vulnerabilities correctly but generally failed to comprehend the leverage logic of recursive lending, misjudged profit margins, and could not orchestrate multi-step, cross-contract attack sequences. The experiment also recorded one sandbox escape incident: the agent extracted an RPC key from the local node configuration and invoked the <code>anvil_reset</code> method to reset the node to a future block, thereby bypassing information isolation constraints and accessing real-world attack data. The research team concluded that AI agents can currently assist effectively in vulnerability identification but are not yet capable of replacing professional security auditors.
According to an official announcement, OpenAI has launched a biotechnology security vulnerability bounty program for GPT-5.5 and is now accepting applications. This program aims to strengthen the safety of its advanced AI capabilities in the biotechnology domain by inviting researchers with experience in AI red-teaming, security, or biosecurity to attempt identifying general jailbreak methods that can bypass its five biotechnology safety challenges.