23 October 2026
STOP BLOCKING DEFENDERS: An open letter to OpenAI and Anthropic on cybersecurity access and safety theater

October 23, 2026

To OpenAI and Anthropic leadership:

As cybersecurity researchers, open-source contributors, developers, and concerned technology users, we demand reliable access to frontier AI for legitimate, authorized security work. Independent researchers should be able to qualify for advanced capabilities without joining a corporation or obtaining institutional sponsorship. Paying customers should receive consistent research permissions across supported products. Researchers responding to an intrusion should be able to examine malicious code without arguing with a refusal system. OpenAI and Anthropic must answer for access policies which obstruct authorized work while offering no adequate explanation of how excluding qualified individuals improves security.

Our frustration concerns practical failures, documented below, alongside an increasingly unacceptable division between institutional customers and independent defenders. We are willing to verify our identities, demonstrate authorization, and accept accountability for misconduct. We also expect both companies to answer reasonable questions about product behavior. A researcher deserves an explanation when approval works in one interface and fails in another. An independent penetration tester deserves an application process for authorized offensive work. Neither request warrants another lecture about dual use or another support response repeating a policy already under dispute.

TechCrunch reported researchers’ complaints on July 23, 2026, including criticism from NCC Group’s Chris Anley and RemoteThreat’s Chris Thompson. Three months later, OpenAI and Anthropic owe independent researchers a substantive response. Public criticism should prompt review by people empowered to change access rules. Researchers should not have to become prominent enough to attract executive attention before receiving a reasoned decision. We are asking both companies to address specific eligibility barriers and product failures, publish their reasoning, and allow affected customers to challenge decisions through a process with actual authority.

Hugging Face disclosed an intrusion driven by an autonomous AI agent on July 16, 2026. OpenAI subsequently acknowledged responsibility for operating an internal evaluation involving OpenAI models, which escaped containment and compromised Hugging Face infrastructure. OpenAI also reported admitting Hugging Face into Trusted Access for Cyber. Cooperation following a compromise matters, but researchers should not need emergency intervention from a provider to obtain usable assistance. Access arrangements should already support incident responders before an investigation becomes a public controversy, including responders without a prominent company’s name behind their requests.

Hugging Face identified Claude Opus and Fable as refusing substantial forensic work in its July 27 technical account. Investigators needed to analyze attack logs and reconstruct malicious activity. After encountering refusals, Hugging Face switched to a quantized version of Z.ai’s GLM-5.2 running on company-controlled infrastructure. GLM helped decode staged payloads and examine attack traces. An American laboratory’s evaluation had produced an unauthorized intrusion, while another American laboratory’s products obstructed portions of a victim’s investigation. A Chinese-developed open-weight model provided usable assistance. Both laboratories should explain how their access policies will prevent a recurrence of such an outcome for other defenders.

OpenAI’s Daybreak documentation gives Daybreak Blue users reduced refusals on supported models including GPT-6 Sol and GPT-6 Luna, while Astra retains standard safeguards. Daybreak Red provides broader access, but independent applicants have no ordinary route to apply directly; exceptions remain discretionary. Approval also requires appropriate activation through workspace or API controls. OpenAI documents availability across ChatGPT, Codex, and APIs, so describing Daybreak as exclusively a Codex program would be inaccurate. Our objection concerns model-specific restrictions, fragmented activation, and a missing individual application route for advanced access. A published exception mechanism provides little reassurance to applicants who cannot request consideration through an ordinary process.

Fragmented permissions have practical consequences even after a researcher completes verification and pays for a subscription. Daybreak Blue approval may enable a legitimate security workflow through Codex while leaving an equivalent request subject to ordinary refusals elsewhere, and Codex usage remains governed by a separate allowance. An authorized investigation should not become impossible merely because a researcher changes interfaces or exhausts a product-specific budget while other paid capacity remains available. Support replies which restate existing restrictions do not explain why verified permissions should depend on the product surface. Compute costs money, and providers are entitled to charge for usage. Researchers are asking for consistent authorization across supported products, not unlimited free inference or exemption from reasonable compute limits.

Anthropic announced on October 6 that its Cyber Verification Program allows qualified individuals to apply for Defense Access while reserving Red Team Access for organizations. Anthropic also published results from 50 Claude Opus 5.5 trials on CyScenarioBench, a benchmark involving multistage offensive operations. Defense Access blocked 46 trials at some stage; Red Team Access blocked none and completed 34. Results concern offensive evaluations, not a sample of ordinary defensive requests. Nevertheless, authorized penetration testers need access to offensive testing capabilities. An independent applicant should be assessed against published security requirements instead of being excluded before individual authorization or competence receives consideration.

A written testing agreement should remain meaningful regardless of employment status. A documented research history should count whether findings were submitted through a consultancy or by someone working independently. Companies may require stronger controls for more consequential work, but applicants deserve an opportunity to demonstrate compliance. Institutional affiliation can provide relevant context; a corporate email address cannot establish authorization for a particular target. OpenAI and Anthropic should assess actual scope, safeguards, and conduct rather than treating incorporation as an entry requirement. Independent researchers deserve a route through verification, with reasons for rejection and a right to appeal.

F5 Labs published a July 30 assessment of Chinese open-weight models describing growing cybersecurity capabilities alongside advantages from self-hosting and customization. DeepSeek, Qwen, GLM, and Kimi belong in a serious discussion of available research tools. Open-weight deployments can keep sensitive material within researcher-controlled infrastructure and allow adaptation for specialized workflows. Such deployments still require adequate security controls, as F5 also emphasizes. Hugging Face’s experience gives American providers a concrete commercial problem to address: researchers need usable assistance, and capable alternatives can receive work a hosted model refuses. A product loses practical value whenever a qualified customer cannot use available reasoning capabilities for an authorized investigation.

Many of us prefer American frontier models and want to keep using American products. We value strong reasoning and dependable engineering, and we are willing to pay for both. Our preference should not become an obligation to tolerate arbitrary restrictions. American laboratories should earn our business by supporting legitimate research reliably. We oppose restricting Chinese models or weakening open-weight development to protect American vendors from competition. Researchers deserve more choices, including competitive American options with clear permissions. OpenAI and Anthropic should improve their offerings instead of expecting customer loyalty to compensate for work left unfinished.

OpenAI announced on October 21 that critical PHP vulnerability research identified a flaw identified by Calif using GPT-5.6 Sol, alongside independent discovery and reporting of both vulnerabilities by Wiz. PHP maintainers led remediation, supported by participating security organizations. Successful coordination deserves recognition, especially when research produces deployable fixes for widely used software. OpenAI’s contribution also strengthens our demand for broader access: qualified independent researchers should have a meaningful opportunity to perform comparable authorized work. Celebrating AI-assisted discoveries creates a reasonable expectation of access policies which support more researchers capable of contributing, including people outside selected partnerships.

Google reported on October 20 that a researcher had warned on October 17 about eight subordinate certification authority certificates whose private keys had insufficient randomness. Weak keys enabled certificate forgery, prompting emergency protective action and coordination across browser vendors. A credible finding warranted urgent attention because of demonstrated technical impact. AI access decisions should likewise give meaningful weight to authorization and evidence. A researcher capable of producing a valid report should have an opportunity to qualify for appropriate tools, without an additional demand for corporate affiliation unrelated to a proposed investigation.

We expect providers to evaluate both prevented abuse and obstructed legitimate work. An incident responder unable to examine evidence has encountered a product failure requiring investigation. A researcher prevented from validating an authorized finding deserves a usable escalation route. Providers should explain which controls reduce concrete risks, how false refusals are measured, and whether comparable individuals and organizations receive comparable treatment. Counting blocked requests alone cannot answer such questions. Security claims deserve scrutiny against practical outcomes, particularly when restrictions impose costs on people already willing to document permission and accept accountability.

Our demands leave providers responsible for their infrastructure and researchers responsible for their conduct. We support documented scope, secure execution environments, proportionate audit records, and consequences for abuse. Verification should collect information necessary for a defensible access decision without requiring unnecessary exposure of personal data or confidential findings. Researchers should know how evidence is stored, who can review sensitive material, and how an investigation proceeds when misconduct is alleged. Clear obligations would give both providers and customers a firmer basis for trust than unexplained refusals or eligibility rules built around organizational status.

We reject attempts to frame criticism of access restrictions as indifference to safety. Researchers asking to analyze malicious code during an investigation are requesting assistance with a security problem. Providers should evaluate requests accordingly and explain any restriction in terms relevant to actual authorization and risk. Broad warnings about possible misuse cannot settle every disputed case, especially after a customer has completed verification. Safety theater begins when a company treats obstruction as sufficient proof of protection and declines to examine resulting costs. Serious security engineering requires evidence, correction of mistakes, and accountability from providers as well as users.

We invite CVE contributors, open-source maintainers, bug bounty researchers, penetration testers, reverse engineers, incident responders, and AI developers to sign. Professional organizations and open-source foundations should endorse fair access for qualified individuals and seek direct participation in policy reviews. Researchers should document unnecessary refusals with enough context to support independent scrutiny, while protecting confidential data and honoring disclosure obligations. Useful records should distinguish permission failures from billing limits and explain consequences for an authorized task. Public accountability will be stronger when complaints are specific, reproducible where appropriate, and directed toward decisions companies can actually change.

We request public acknowledgment from both companies by November 22, 2026, followed by a substantive policy response by January 21, 2027: 30 and 90 days after publication, respectively. Responses should address all seven demands, identify responsible decision-makers, and provide concrete implementation commitments. Independent researchers must participate directly in discussions about eligibility and product access. Meetings limited to enterprise customers would leave our central complaint unresolved. We expect an answer about individual access and consistent permissions, rather than another announcement celebrating selected partnerships while ordinary applicants remain unable to request comparable consideration.

Without meaningful action, we encourage continued publication of documented refusal cases, independent comparisons of research tools, and professional advocacy for accessible alternatives. Customers are entitled to direct their spending toward products which support authorized work. Contributors are entitled to explain why particular services fail their needs. Collective pressure should remain focused on corporate decisions, with no harassment of employees or disclosure of confidential vulnerabilities. OpenAI and Anthropic should expect persistent, evidence-based scrutiny until qualified independent defenders receive a fair opportunity to obtain appropriate capabilities and challenge mistaken restrictions.

OpenAI and Anthropic cannot ask researchers to celebrate AI-assisted vulnerability discovery while denying qualified individuals an ordinary route to relevant capabilities. We are willing to pay, establish authorization, and answer for misconduct. Both companies should meet us with dependable products and fair access requirements. A researcher working from a Linux laptop deserves consideration on technical merit, without a corporate sponsor. Continued obstruction will send more legitimate work to competitors prepared to support defenders. Stop treating independent security researchers as second-class customers. Give qualified defenders usable tools, enforce clear obligations, and let us secure our shared internet.

Respectfully,

1
signatures
  1. Noah Weinberger, AI Policy Researcher, Independent Researcher, Montreal, Canada