Before you ask an AI to audit your code: is your account at risk?

PHP shipped a critical security advisory on October 21, and since then AI code audit prompts have been making the rounds on X. Using one makes sense: a model can go through in minutes what takes a person an afternoon.

The assumption underneath is that if the code is yours, asking a model to review it carries no risk. Two things that surfaced over the weekend test that assumption without breaking it. An OpenAI employee said blocks on reverse engineering requests in Codex were mostly accidental. And a well-known user got a warning from Google about his Gemini usage, which he suspects came from auditing his own code.

One thing up front: neither case is a new policy, and nobody here reports a banned account. Below I keep what's documented apart from what's a guess or an opinion.

Update, Oct 26

At 1:20 AM ET on October 26, Varun Mohan, who leads the Antigravity team at Google DeepMind according to Google's I/O 2026 speaker page, replied to Artem on X: "Let's DM, the email is extremely unclear but Antigravity should not be banning you from any service other than AGY given misuse". In other words, if there was misuse in Antigravity, the consequence should stay inside Antigravity.

At 4:41 AM ET, Artem said two Googlers had reached out and that they'd soon get to the bottom of what happened and whether Antigravity was involved. Digit wrote the story up with the same facts and notes that it's still unclear what triggered the warning.

That changes the tone, not the substance. Mohan is an employee posting on X. It isn't a Google statement or a published policy. He doesn't say what triggered the warning or confirm it was Antigravity, and that's exactly what's still unknown. What's new is that the person running the product publicly calls the email unclear and says misuse in Antigravity shouldn't spill over into the rest of your Google services.

What an OpenAI employee said about Codex and reverse engineering

On October 25 at 7:46 PM ET (20:46 in Argentina), Eric S. Raymond complained on X that OpenAI and Anthropic were blocking people who use their models to strip spyware out of smart TVs, work that requires reverse engineering. He framed it himself as something he'd heard secondhand.

Under an hour later, roon, who works at OpenAI, replied: "these blocks are mostly accidental and a result of difficult decision boundaries; there is no prohibition against reverse engineering with codex. it's getting fixed".

That's encouraging, but it's a personal post on X. It isn't a published policy or a company statement, and it doesn't say what's being fixed or when.

What is published is OpenAI's usage policies. The version updated on October 21 prohibits the destruction, compromise, or breach of another's system or property, including malicious or abusive cyber activity, as well as unsolicited safety testing. The only mention of reverse engineering I found is about extracting the protected reasoning of OpenAI's own models, not about analyzing your devices or your code. The operative word is "another's". Auditing your own stuff isn't prohibited, but an automated filter can't always tell it's yours.

What happened to one Gemini user

Artem Russakovskii, founder of Android Police and APKMirror, posted that he'd received several notifications from Google. Per the text he shared, the notice read "Service: Gemini" and gave the reason as "Trying to interfere with or disrupt Google's services or infrastructure, or attempting to use Google services for disrupting others' infrastructure or services". It also said his access hadn't been restricted, but that he could lose it if he continued.

Two details matter. He says he was driving and not using Gemini when the notices arrived, so this wasn't a live reaction to something he was doing. And the email didn't say what triggered it. In a follow-up he said Google's own page confirmed the notice was legitimate but showed no details.

Later that evening he shared his best guess. The day before, he'd installed Google Antigravity for the first time to run a security analysis of his own code, using version 2 of the Black October Prompt, an audit prompt circulating on X that builds on Cloudflare's security-audit-skill. He ran it with Gemini 4 Argon, which he says ran out of credits and killed the task. He presents it as the only guess he has.

One reply came from another user who said he got the same warning, and that the only security topic he'd asked about was the PHP2Shell CVE. That's a second account from the person involved, not a pattern.

As of the morning of October 26, Google hadn't issued a statement. What did happen is a reply from an employee, covered in the update above.

What Google's policy says

Google's Generative AI Prohibited Use Policy asks you not to compromise the security of Google's or others' services, and lists "interference with, or disruption to Google's or others' infrastructure or services". The reason on Artem's notice uses almost the same words. The same page says exceptions can be made for educational, documentary, scientific, or artistic reasons, but not how those are evaluated.

The nuance X skipped

On X the story got compressed into "Codex allows reverse engineering, Gemini bans you." Neither half is proven. roon isn't speaking for a written policy, and neither is Varun Mohan. Artem wasn't banned. He got a warning, and the link to his audit is his own suspicion.

In the same thread, TokenGremlin said Claude is "still a gray area" and will assume you're a state hacker going after a power grid. That's a joke-flavored opinion, not a report.

Timeline (US Eastern Time)

• Oct 21: php.net posts a critical security advisory covering two remote code execution flaws and asks everyone to update to 8.5.12, 8.4.27, 8.3.36, or 8.2.35.
• Oct 23: version 2 of the Black October Prompt starts circulating. The same day php.net adds an informational advisory about a separate issue.
• Oct 25, 7:46 PM: Eric S. Raymond complains about blocks at OpenAI and Anthropic.
• Oct 25, 8:28 PM: Artem Russakovskii posts the Gemini warning.
• Oct 25, 8:43 PM: roon replies that reverse engineering with Codex isn't prohibited.
• Oct 25, 9:05 PM: Artem shares his theory about the Antigravity audit.
• Oct 25, 10:39 PM: another user says he got the same warning.
• Oct 26, 1:20 AM: Varun Mohan of Antigravity replies that the email is extremely unclear.
• Oct 26, 4:41 AM: Artem says two Googlers have reached out.

Proven, alleged, rumor

Proven (a document or original post shows it):

• An OpenAI employee wrote that reverse engineering with Codex isn't prohibited and that the blocks are being fixed. What's proven is that he wrote it, not that it's policy.
• OpenAI's usage policies prohibit attacking other people's systems and unsolicited safety testing.
• Google's policy prohibits interfering with its own or others' infrastructure, in the same language as Artem's notice.
• Artem got a Gemini warning with no details and no restriction on his access.
• Varun Mohan, who leads Antigravity, wrote on X that the email is extremely unclear and that Antigravity shouldn't ban you from other services over misuse. What's proven is that he wrote it, not that it's Google policy.
• PHP disclosed critical vulnerabilities on October 21, with patched versions.

Alleged (someone says so, unconfirmed):

• That the warning came from using Gemini to audit his own code. That's Artem's theory.
• That another user got the same warning for asking about PHP2Shell. One testimony.
• That the Codex blocks were accidental. An employee says so, not the company.
• That two Googlers reached out to Artem. He says so.
• That misuse in Antigravity only affects Antigravity. The product lead says so on X; no published policy states it.

Rumor or opinion:

• That Gemini bans accounts for code audits. These threads don't contain a single banned account.
• That Claude treats any audit as an attack. An opinion.
• That OpenAI and Anthropic block people from cleaning up TVs. Raymond described it as secondhand.

What I'd do before asking an AI to audit your code

This isn't legal advice. It's what makes sense given what's published today.

1. Use an account that isn't the one holding your email and files. If a warning ever turns into a locked account, it shouldn't take anything important with it. The user who replied to Artem landed on the same idea.
2. Only audit code and devices you own or have written permission to test. Both companies' policies draw the line at other people's systems and unsolicited testing.
3. Put that context in the prompt: which repository it is and that it's yours. It guarantees nothing, but it gives the model, and any human reviewer, the missing piece.
4. Ask for findings and fixes, not ready-to-run exploits against servers you don't control.
5. Don't paste keys or customer data into the prompt. An audit doesn't need them.
6. Save the exact prompt and note which model you used and when. If a warning shows up with no details, you can reconstruct what happened and appeal with facts. OpenAI's policies say you can appeal if you think they got an enforcement decision wrong.
7. If a Gemini filter starts refusing every prompt, a moderator with a Google badge on the Google AI Developers Forum recommended starting a new chat, stating in the prompt that you're doing defensive security research, and reporting the false positive through the feedback link on the block message. That reply is from September and covers prompts blocked inside a chat, not account warnings.
8. If PHP is your problem, AI isn't step one. Update to the patched versions from the October 21 advisory.

Bottom line

Running an AI code audit is still a good idea, and nothing here says otherwise. What these cases show is that automated filters can't tell whether the code is yours, and that warnings can arrive with no explanation. With a separate account and a record of what you asked, the worst case shrinks from a loss to an annoyance.

FAQ

Does OpenAI prohibit reverse engineering with Codex?

An OpenAI employee wrote on X that it doesn't, and that the blocks were mostly accidental. That isn't published policy. The written usage policies prohibit attacking other people's systems but say nothing about reverse engineering your own devices.

Does Google ban accounts for auditing code with Gemini?

There's no evidence of that. One user got a Gemini account warning without any access restriction, and another says he got the same one. The link to code auditing is the affected user's own theory. Google hasn't issued a statement; the Antigravity lead replied on X that the email is extremely unclear and that Antigravity shouldn't ban you from other services over misuse, but that isn't published policy.

What does the Gemini account warning mean?

Based on the text Artem shared, Google flagged an attempt to interfere with its own or others' services. The notice doesn't say what triggered it and states that access wasn't restricted. What triggered it is still unknown; Artem says two Googlers are looking into it with him.

Should I use a separate account for AI code audits?

It's a cheap precaution. If something goes wrong, the problem stays in an account that doesn't hold your email or your files.

What should I do if I run a PHP site?

Update to 8.5.12, 8.4.27, 8.3.36, or 8.2.35 depending on your branch, as the October 21 php.net advisory asks. An AI audit can come after that.

Sources

• roon (OpenAI employee) on X, Oct 25
• Eric S. Raymond on X, Oct 25
• Artem Russakovskii: the Gemini warning
• Artem Russakovskii: the notice is real, no details
• Artem Russakovskii: his theory about the audit
• Reply from another user with the same warning
• TokenGremlin on Claude
• Black October Prompt v2
• Cloudflare security-audit-skill on GitHub
• OpenAI usage policies
• Google Generative AI Prohibited Use Policy
• php.net: October 21 security advisory
• php.net: October 23 informational advisory
• Varun Mohan (Antigravity) on X, Oct 26
• Google I/O 2026 speakers, with Varun Mohan's role
• Artem Russakovskii: two Googlers reached out
• Digit's coverage of the case
• Google AI Developers Forum: false positives in Gemini filters