Security teams have spent the past two years discovering an uncomfortable irony: the models that can find vulnerabilities at scale are the same models vendors are most cautious about handing to the people who fix them. A newly published open letter addressed to OpenAI and Anthropic — hosted on openletter.earth under the banner "Stop blocking defenders" — makes that irony its central charge, arguing that current access controls protect reputations more than they protect infrastructure.
The core charge: guardrails aimed at the wrong customer
The letter's argument, stripped to its skeleton, is about targeting. Frontier labs build abuse-prevention systems calibrated against the worst-case actor: the person who wants working exploit code, malware variants, or a shortcut past a patch. But those systems classify by behavior and topic rather than by intent or accountability, so a red teamer running an authorized assessment and an attacker running the same prompt land in the same bucket. The result is that the defensive side of the security industry — pen testers, detection engineers, incident responders, vulnerability researchers — absorbs the friction, while the adversary, who has fewer qualms about jailbreaking, routes around it.

Figure: HEXOPS original illustration inspired by the topic (not a copy of the source article imagery) · Credit / further reading: openletter.earth
The phrase the letter leans on is "safety theater": the claim that visible refusals and blocked requests manufacture the appearance of governance without improving real-world security outcomes, because the marginal attacker is not the one being stopped. Whether or not you accept the framing wholesale, the underlying observation is testable, and plenty of practitioners have their own anecdotes to match it.
Why builders should care
The false-positive tax
If you ship anything that touches security workflows — SIEM enrichment, log triage, detection-as-code generation, threat-intel summarization — you have probably already built a fallback layer without naming it that way. Teams route sensitive prompts through self-hosted open-weight models, keep a second tab open for cases the frontier API refuses, or pre-process inputs until they slip under the classifier. That is engineering time spent on workarounds instead of product. It also creates a quiet split in your architecture: the capability you advertise and the capability you can actually deliver inside a customer's environment diverge, and the divergence stays invisible until someone hits it during an evaluation.
Procurement and liability
Enterprise buyers now ask vendors to document which models they use and how refusals are handled. A security product that stalls mid-incident because an API declined a request is not hypothetical — it is a support ticket with a clock on it. The letter effectively asks labs to publish clear, appealable criteria for security-related access, so downstream vendors can build against a stable contract instead of reverse-engineering opaque policy shifts. Absent that, the safe design choice is to over-provision local compute, which favors well-funded teams and consolidates the market around whoever can afford it.

Figure: HEXOPS original illustration inspired by the topic (not a copy of the source article imagery) · Credit / further reading: openletter.earth
The vendors' side of the ledger
It is worth being fair to the labs. The abuse surface is genuinely ugly: model-assisted biosecurity, autonomous exploitation, and scaled social engineering are real risks, and a vendor that loosens access after an incident pays for it publicly. Verification is also hard at scale — the gap between "verified security researcher" and "person who claims to be one" is precisely the gap attackers exploit, and nobody has shipped a clean solution to identity assurance for API access. Skeptics will also note that open letters are cheap and that the signatory mix matters; readers should check the current list and the specific asks at the source rather than treating any summary as settled.
The stronger version of the ask, then, is not "remove the guardrails" but "make them auditable and appealable." A tiered access program with documented review, an appeals path with a response commitment, and a public postmortem when a legitimate defender gets blocked would cost the labs far less than they fear — and would give the ecosystem something concrete to build against.
What to watch next
Three signals matter. First, whether either lab responds publicly or quietly revises its trust-and-safety documentation; silence from one and a policy-page edit from the other is itself informative. Second, whether a third-party identity or attestation layer for security access gains traction — that is the piece both sides say they want and neither has built. Third, whether enterprise buyers start writing refusal-handling into RFPs, which would convert a policy argument into a purchasing one. Historically, purchasing is the faster lever.
Key takeaways
- An open letter to OpenAI and Anthropic argues that abuse-prevention controls block legitimate defenders while doing little against determined attackers.
- The criticism is about targeting, not existence: behavior-based classifiers cannot reliably separate authorized red teams from adversaries.
- Builders already pay a false-positive tax in workarounds, split architectures, and failed customer evaluations.
- The practical ask is auditable, appealable, tiered access — not an open door.
- Identity verification at scale remains the unsolved piece on both sides of the argument.
Next steps
- Read the letter and its current signatory list directly, and check whether the specific asks map to your use case.
- Audit your stack for hidden frontier-API dependencies in security-sensitive paths, and document the fallbacks you already use.
- If you buy security tooling, add refusal-handling and model-substitution questions to your evaluation checklist.
- Track trust-and-safety documentation changes at both labs over the next quarter, and note any published appeals process.
Sources
- Open letter, "Stop blocking defenders: an open letter to OpenAI and Anthropic on cybersecurity access and safety theater" — https://openletter.earth/stop-blocking-defenders-an-open-letter-to-openai-and-anthropic-on-cybersecurity-access-and-safety-theater-e653e5cd



