San Francisco-based AI cybersecurity startup Hacktron identified vulnerabilities in OpenAI’s systems during authorized security research using Anthropic’s Claude. Led by co-founder and CEO Zayne Zhang, the team reported its findings to OpenAI and received a $6,500 bug bounty, giving the young company a concrete business milestone from its cybersecurity research.
Key Takeaways
- Hacktron, a San Francisco-based AI cybersecurity startup led by Zayne Zhang, identified vulnerabilities in OpenAI’s systems.
- The company used Anthropic’s Claude during authorized security research to test the vulnerabilities.
- Researchers accessed an OpenAI employee account and a connected Codex account but stopped before accessing internal code.
- Hacktron reported the findings to OpenAI through its security reporting process.
- OpenAI paid Hacktron a $6,500 bug bounty and revoked affected tokens and sessions.
Hacktron Identifies Vulnerabilities in OpenAI Systems
Hacktron identified security gaps in OpenAI’s infrastructure while investigating vulnerabilities that could potentially affect ChatGPT and Codex accounts. The research was conducted as part of the company’s work examining security weaknesses at artificial intelligence companies.
The startup’s research focused on OpenAI’s community help forum and the relationship between access to that forum and connected OpenAI accounts. Hacktron reported that a vulnerability could allow an account associated with the community platform to be compromised.
The company’s approach fits a broader pattern of founders using AI tools to operate with lean teams.
The research involved Hacktron’s security team and was conducted under an authorized security research framework. The company reported its findings to OpenAI rather than retaining access after confirming the vulnerability.
Zayne Zhang, Hacktron’s co-founder and CEO, has described the company as a cybersecurity startup focused on examining vulnerabilities that could be exploited by AI agents. The company is based in San Francisco and has fewer than 10 employees, according to reporting published about the research.
The company’s size provides important context for the business story. Hacktron was launched less than a year before the disclosure, meaning the security research represents one of the startup’s early publicly reported milestones.
The company’s work also involved two different technology companies. Hacktron used Anthropic’s Claude to assist its research against OpenAI’s systems, making the project an example of one AI company’s technology being used in authorized security testing involving another AI company.
Claude Assists Hacktron’s Authorized Security Research
Hacktron used Claude during its investigation after gaining access to Anthropic’s Cyber Verification Program. Zhang said the program relaxed certain cybersecurity restrictions on Claude for authorized security research.
The team used Claude as part of its effort to test whether the identified vulnerability could be exploited. The research was not limited to identifying a theoretical security weakness; the researchers tested the vulnerability against OpenAI’s systems and established that the access path could affect an OpenAI employee account.
The use of Claude was therefore part of a controlled security investigation rather than an independent attempt to attack OpenAI without authorization. Hacktron was participating in OpenAI’s bug bounty process, which provides a framework for researchers to report qualifying security vulnerabilities.
The research also shows the role AI tools can play in the work of a small cybersecurity team. Hacktron used Claude during technical security research while its researchers directed the investigation and determined how to proceed. The company’s disclosure and subsequent reporting identified the work as authorized vulnerability research.
Hacktron’s use of Claude is particularly relevant to the company’s business model because the startup specializes in AI cybersecurity. Its research combines security expertise with artificial intelligence tools to examine vulnerabilities affecting AI systems and their connected services.
The company did not present Claude as the sole actor behind the research. The investigation was conducted by Hacktron’s research team, with Claude used as a tool during the authorized testing process. That distinction is central to understanding the reported work.
Other small AI companies have also built businesses around lean teams and artificial intelligence. For example, bootstrapped AI startup growth has been reported in connection with founders using AI to support operations and product development.
Researchers Access an OpenAI Employee Account
Hacktron’s testing progressed beyond identifying a possible vulnerability. The team was able to access an OpenAI employee’s account and use the employee’s Codex account to suggest changes to an internal code repository.
Hacktron said the researchers stopped before accessing internal code. The team instead used the access to establish the potential impact of the vulnerability and then reported the issue to OpenAI.
The reported access demonstrated that the security gap extended beyond the community forum itself. The researchers were able to connect the vulnerability to OpenAI accounts and then use the affected Codex account to interact with an internal development environment.
The reported access also explains why the discovery qualified for attention under OpenAI’s security reporting process. Rather than identifying only a low-impact configuration problem, Hacktron established that the vulnerability could create access to an employee account and connected development tools.
Hacktron’s decision to stop after demonstrating the access and to report the findings limited the research to the purpose of establishing the vulnerability. The company said it did not access internal code during the test.
The incident involved systems operated by different organizations, including the community platform used by OpenAI and OpenAI’s own account infrastructure. Reporting on the incident said the researchers identified vulnerabilities involving the connection between those systems.
For Hacktron, the research provided a concrete example of the type of security work the startup was established to perform. The company investigated a vulnerability affecting an AI company, tested its potential impact under an authorized framework and documented the result for the affected organization.
Hacktron Reports the Findings to OpenAI
Hacktron reported the vulnerability to OpenAI after confirming the account-access issue. The company’s disclosure process allowed OpenAI to review the findings and make changes to its systems.
OpenAI said it narrowed permissions on Community sign-in tokens and revoked affected tokens and sessions after receiving the researchers’ findings. Those actions addressed the access mechanism identified during Hacktron’s investigation.
The reporting process also separated the security research from unauthorized exploitation. Hacktron was operating within a bug bounty framework and disclosed the vulnerability to the affected company after establishing its impact.
The company’s disclosure became public after the research was completed and the vulnerability had been addressed. The public account of the incident identified Hacktron’s research team, Zhang’s role as co-founder and CEO, the use of Claude and the subsequent bug bounty payment.
The episode also gave Hacktron a public example of its cybersecurity work involving a major AI company. The startup’s research was tied directly to a documented security finding, a responsible disclosure and a payment from the affected company.
For a small company, the sequence connects technical research with a commercial outcome. Hacktron investigated a vulnerability, participated in a formal bug bounty process, reported the finding and received compensation after OpenAI addressed the issue.
The sequence also resembles the operating model seen among other small AI companies, where small startup teams using AI have been built around specialized technical work without large initial workforces.
OpenAI Responds and Pays a $6,500 Bug Bounty
OpenAI paid Hacktron a $6,500 bug bounty following the company’s disclosure. Hacktron identified the payment as compensation for its security research and discovery.
The payment gives the startup a direct financial outcome from its technical research. Unlike a funding announcement, the bounty was tied to a specific cybersecurity finding and the process used to report that finding to OpenAI.
OpenAI also confirmed that it responded to the research. The company said it narrowed permissions on Community sign-in tokens and revoked affected tokens and sessions after receiving Hacktron’s report.
The response established a sequence from discovery to remediation: Hacktron identified the vulnerability, tested its potential impact, reported the findings, and OpenAI changed the affected access controls.
Hacktron’s work also illustrates the operating model behind a specialized cybersecurity startup. The company used a small research team, an AI tool and formal vulnerability-reporting channels to investigate a security problem involving a large technology company.
The company’s reported size adds another specific detail to the business story. Hacktron has fewer than 10 employees and had been operating for less than a year when the research became public.
The $6,500 payment followed the technical work rather than preceding it. Hacktron therefore turned a research activity into a documented business result through the bug bounty process.
For the startup, the episode links its stated focus on AI security with a public security finding involving OpenAI. The research also established the company’s ability to identify and responsibly report a vulnerability affecting accounts and connected development tools.
Frequently Asked Questions
What is Hacktron?
Hacktron is a San Francisco-based AI cybersecurity startup led by co-founder and CEO Zayne Zhang. The company focuses on security research involving artificial intelligence systems.
How did Hacktron identify OpenAI vulnerabilities?
Hacktron investigated OpenAI’s infrastructure and identified security gaps involving its community help forum and connected accounts. The team then used Anthropic’s Claude during authorized testing to establish the vulnerability’s potential impact.
How was Claude used in Hacktron’s security research?
Hacktron used Claude as part of its authorized security testing after gaining access to Anthropic’s Cyber Verification Program. The team used the model during its investigation while researchers directed the security work.
How much did OpenAI pay Hacktron?
OpenAI paid Hacktron a $6,500 bug bounty after the startup reported its findings. The payment followed the company’s vulnerability research and disclosure process.
What changes did OpenAI make after the findings?
OpenAI said it narrowed permissions on Community sign-in tokens and revoked affected tokens and sessions after receiving Hacktron’s report.
