[BUG] CVP approved org still blocked by cyber safeguards — org ID confirmed matching, appeal form shows no fields

Even I am facing the same issue since yesterday

Yep, same issue, cant even continue past sessions which really sucks.

same issue

Exactly the same here. Claude Opus 5 refused to answer any of my questions, and it wouldn’t even explain what cybersecurity is when I asked.
I’m ready to cancel my subscription.

I’ve switched to ChatGPT now...

Yeah at this point I can't help but wonder why I don't just pay for openai or even chinese models without any of this bullshit.

same issue. no sessions are working T_T

so what is the soultion for this?

Since the day I added it, I’ve confirmed my CVP, and now I’m sitting here and can’t use Claude... I really hope the situation will be resolved.

Want my refund back

same problem can't even do basic crap, super annoying. switched over to codex (which sucks even more but at least it doesn't fail on basic tasks)

CVP doesnt do anything. It's sad they neutered every model.

On Sun, Aug 16, 2026 at 8:07 AM D. @.***> wrote:

totekuh left a comment (anthropics/claude-code#84689) <https://github.com/anthropics/claude-code/issues/84689#issuecomment-5308088698> same problem can't even do basic crap, super annoying. switched over to codex (which sucks even more but at least it doesn't fail on basic tasks) — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/84689?email_source=notifications&email_token=BC7QLMZNDLQVQCN3JMBSFML5KHE27A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZQHAYDQOBWHE4KM4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-5308088698>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BC7QLM5J2IPEAQ6EUQ6DKUD5KHE27AVCNFSNUABFKJSXA33TNF2G64TZHM4TGNZSGUZTINZVHNEXG43VMU5TKMBYGYZDCNZSGA32C5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/BC7QLMZWAYQBSC2CFS5LMZ35KHE27A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZQHAYDQOBWHE4KM4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/BC7QLMY7CT7QPCAW4ICTXMD5KHE27A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZQHAYDQOBWHE4KM4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you commented.Message ID: @.***>

Same here, cvp seems to only work on 4.8 since like a week or two.... and even on 4.8 i am getting more cyber refusals now...

Same here, you guys have any suggestion to replace claude code?

Same here, you guys have any suggestion to replace claude code?

Give GPT/Sol a try, its "cyber filter" is by default much less erratic even without any apparently useless "CVP", that is what I settled on after cancelling claude. Whether claude is superior does not matter if I am unable to use even Opus 5 for unknown random "cyber reasons" because Anthrophic silently changed something two weeks ago.

The newly released GLM 5.3 works as well, it is likely as smart as opus 4.8, just lacks the vast knowledge. But that is still better than $200 for a useless tool that refuses to fix bugs because it's in a cyber mood today!

I've now run the same workflow across DeepSeek V4 Pro, Opus 5, and GPT 5.6 Sol (Codex) to compare

Claude is still the most creative of the three for this kind of work. I can't put a hard number on it, but it feels like Claude has less internal pushback against running for instance cURL against a target, chaining odd behavior into an actual exploit, or just thinking up ideas to go test. Codex finds fewer bugs the same way — but hand it a list of entrypoints and weird behavior and it's excellent at pushing that into a real vulnerability with a high-quality, real-world-impact writeup. It's a great closer, not as good an opener.

DeepSeek V4 Pro is solid too — it surfaces odd/complex behavior at a similar rate to to Opus 4.6. Not a bad combo is DeepSeek to flag anything weird, then Codex to relentlessly dig into it and write it up. Though this is also pre DeepSeek price increase.

The newly released GLM 5.3 works as well, it is likely as smart as opus 4.8, just lacks the vast knowledge.

@Hoernchen What closes it more than raw model swap is giving the dispatched agent skills/reference docs up front (past CVEs, framework-specific gotchas, auth-bypass patterns, etc). Helps to narrow the gap a lot.

For reference, my general flow:

1. Vulnerability discovery — agent explores entrypoints, flags anomalous/weird behavior
2. Chains — weird behavior gets pushed to a vulnerability, and other bugs/vulns are looked into for potential chains
3. FP checks — kill anything that doesn't hold up before it goes further
4. Report — write up the finding with a strong focus on what the actual impact is to this particular organization.
5. Blind triage and repro — a separate pass (blind to the original findings) re-verifies and reproduces, and takes another look at the supposed security impact.

Another data point, with an additional symptom that I think is more diagnosable than "same here": once a session is flagged, context-free messages are blocked too.

Environment

Org ID:         f076b231-2eeb-43cb-9b5c-cf9b0379d644
Org type:       claude.ai consumer (Pro, stripe_subscription)
Role:           admin
Portal status:  Active — 30-day retention + usage monitoring both enabled
Client:         Claude Code 2.1.251, macOS 25.6.0
Auth:           first-party claude.ai OAuth (no API key)
Models:         Opus 5 -> auto-fallback -> Opus 4.8 -> hard block

Org ID confirmed matching across the CVP approval, the Verification Portal
(program shows Active, not "in review"), and oauthAccount.organizationUuid
in ~/.claude.json.

The Verification Portal account switcher lists two orgs:

CLAUDE CONSOLE   <individual console org>
CLAUDE.AI        <this org>   <- selected, and the one Claude Code runs on

The program page shows Active with the claude.ai org selected, so this is not
the org-mismatch case reported in #84352.

Request IDs

req_011CeYT6a7VqCRYQjD8hJNb5
req_011CeYT8XmurUie4hKTAtQB3

The part that seems worth isolating

After the first block, I switched to Opus 5 and sent a single message with no
technical content at all — a two-word Korean interjection roughly equivalent to
"ugh, dude". That message was flagged:

Opus 5's safeguards flagged this message. ... Switched to Opus 4.8.
Details: [cyber]

and the resulting Opus 4.8 request was then hard-blocked with [cyber]
(req_011CeYT8XmurUie4hKTAtQB3).

So the classifier appears to be scoring accumulated conversation context rather
than the current message. The practical effect is that a flagged session becomes
permanently unusable — no message, however innocuous, will go through — and the
Opus 5 -> Opus 4.8 fallback path inherits the same context and fails identically,
so the fallback provides no recovery.

Work being done

Ordinary Rust development on a C2 framework I own and run locally: workspace
scaffolding, a shared wire-protocol crate, and a web dashboard. No live targets,
no third-party systems.

Expected

With CVP Active on the org the client is authenticated against, dual-use security
development should not hard-block — and in any case a flagged session should not
propagate to unrelated subsequent turns.

Follow-up to my comment above — a correction, and a new data point that I think is
more on-point than my first one.

Correction

In my earlier comment I described the project as "ordinary Rust development on a C2
framework I own and run locally: workspace scaffolding, a shared wire-protocol crate,
and a web dashboard. No live targets, no third-party systems."

That understated it, and I want to correct the record. I wrote it from the repo's
changed-file list without reading the project roadmap. The project is a full C2
framework whose planned scope includes in-memory BOF/COFF execution, .NET
execute-assembly, shellcode injection, SOCKS proxying and port forwarding, a
credential/loot store, malleable-profile-style traffic shaping, and token/privilege
operations. The roadmap scopes it to authorized engagements, CTF, and detection
research in isolated labs — but it is unambiguously offensive security tooling
development, and I should have described it that way.

New block

req_011CeYYtLsPa96NxrApezP4f

Same org, same Active CVP status, environment otherwise unchanged from my
previous comment (Claude Code 2.1.251, macOS, first-party claude.ai OAuth, Pro).

Unlike the block in that comment, this one is not a false positive on innocuous
content. The flagged turn was a roadmap design discussion covering exactly the
capabilities listed above. The classifier fired on genuinely dual-use material.

Why that makes it the more useful case

The help center article classifies "vulnerability exploitation or offensive security
tooling development" as high-risk dual use — blocked by default, and explicitly
described as adjustable through the CVP.

This org's CVP is Active, and that category is still hard-blocked. So either the
approval is not reaching the classifier for first-party claude.ai OAuth sessions, or
"Active" in the Verification Portal does not correspond to an applied adjustment on
the inference path. From the client side the behavior on an Active org is
indistinguishable from having no approval at all.

Two questions that would resolve this for everyone in this thread:

1. Does CVP approval apply to consumer (Pro/Max) claude.ai OAuth sessions in Claude

Code, or in practice only to Console/API-key traffic? The help center lists
Claude.ai and Claude Code as supported surfaces, which is why several of us
applied on consumer orgs in the first place.

1. Is there any way to confirm an approval is actually being applied to a given

request, other than inferring it from whether a block happens?

The Opus 5 -> Opus 4.8 fallback behavior from my earlier comment still reproduces:
the fallback inherits the flagged context and fails identically, so there is no
recovery path within a flagged session.

Same problem here, adding a data point.

I'm a paying Claude Max subscriber, based in the EU (France). My CVP application is stuck in "Under review" on the business verification (KYB) step. Individual identity verification passed fine, but the Persona KYB inquiry link expired, and there is no button anywhere in the portal to restart or regenerate it once the application is in review.

I can confirm the appeal form problem reported here: claude.com/form/cyber-block-false-positive-report-cvp-rejection-appeal loads with no visible fields for me too, so there is effectively no working appeal path.

Channels I've tried, all dead ends: the identity verification help form (submitted twice, no reply in over two weeks), in-app support (loops through the Fin bot and closes the ticket without ever escalating), and sales (no reply). The only thing that reached an actual human so far was emailing privacy@anthropic.com directly.

Org ID: 1925*****************1392c

Could a maintainer route this to the Safeguards/Identity team so my Persona KYB inquiry can be reissued? Related: #84352, #69938.

Same problem here.