What not tell ChatGPT?

The Privacy Trap: Why ChatGPT Isn't as Private as You Think

If you type it into ChatGPT on a Free or Plus plan, assume it's no longer private. Under default settings, your prompts train the model, your history is saved until you manually delete it, and the Memory feature quietly builds a persistent profile of your preferences and facts that carries across every conversation. One consumer comparison even ranked ChatGPT as the least private of four major chatbots by default.

This isn't paranoia — it's documented. A Stanford HAI study analyzed 28 privacy documents from six major U.S. AI companies and found all six train on user chat data by default, some retaining it indefinitely and some allowing human review of transcripts. Lead researcher Jennifer King warns that sensitive information, including files you upload during a conversation, may be collected and used for training.

Even deletion offers less protection than most people assume. During the NYT v. OpenAI litigation, a court order required OpenAI to preserve consumer chats — including ones users had already deleted. Standard deleted chats follow a 30-day deletion schedule that can be extended "for legal or security reasons." Treat anything typed into consumer ChatGPT as potentially permanent.

Here's the trap that catches paying customers: your subscription doesn't buy corporate-grade privacy. As one analysis put it, if you assume your €20-a-month Plus plan delivers the same posture as a corporate data agreement, it doesn't. The tiers break down like this:

• Free and Plus — train on your prompts by default; you must manually opt out
• Pro — the only consumer tier with automatic exclusion of prompts from training
• Business and Enterprise workspaces — OpenAI does not use this content for training by default, with AES-256 encryption at rest

The stakes are real, not theoretical. Samsung engineers leaked confidential semiconductor code through ChatGPT, and a 2024 breach exposed more than 225,000 OpenAI credentials on dark web markets. Meanwhile, 77% of employees paste data into generative AI prompts, and 82% of those pastes come from personal accounts outside company control.

The takeaway for business owners and AI-curious professionals: check which tier you're on before sharing anything sensitive, and route confidential work through a managed Business or Enterprise workspace rather than a personal chat window. The same scrutiny should extend to the tools and agent skills you connect to your AI stack — which is exactly why SkillzBench is building free security scanning and vetting for agent skills, so teams can see what's safe before it ever touches their workflow.

What the Leaks Look Like: Real-World Consequences of Oversharing

It takes just one pasted code snippet to turn a helpful chatbot into a corporate data breach — and the incident reports prove this isn't hypothetical.

The most cited cautionary tale is Samsung. In 2023, engineers pasted confidential semiconductor source code into ChatGPT looking for help with fixes, effectively handing proprietary secrets to a third-party server. According to documented incident reports, 65% of Samsung employees later expressed serious concerns about generative AI security — a hard lesson learned after the leak.

The numbers behind everyday oversharing are staggering. Enterprise security research has flagged over 410 million data loss prevention violations for ChatGPT alone, including attempted shares of Social Security numbers, source code, and medical records. Cyberhaven's tracking shows 34.8% of corporate data entering AI tools is now sensitive — up from 10.7% just two years earlier. And IBM found that 1 in 5 organizations was breached via shadow AI in 2025, adding an average of $670,000 in breach costs.

Credential theft compounds the problem. A 2024 infostealer breach exposed more than 225,000 OpenAI credentials on dark web markets, meaning attackers could log into real accounts and read real conversation histories.

• 410 million DLP violations flagged for ChatGPT, per Zscaler ThreatLabz
• 34.8% of corporate data entering AI tools is sensitive (Cyberhaven)
• 1 in 5 organizations breached via shadow AI in 2025, adding +$670,000 in costs (IBM)
• 225,000+ OpenAI credentials exposed on dark web markets in 2024

What makes this risk insidious is inference. A Stanford HAI study found that even innocuous, health-adjacent queries — say, asking for low-sugar or heart-friendly recipes — can lead algorithms to classify you as "health-vulnerable." That classification then cascades into medication advertising and could theoretically reach insurance companies. Multiproduct companies like Google, Meta, and Microsoft routinely combine chatbot interactions with search queries, purchases, and social media data, so one "harmless" question can enrich a profile that follows you across their entire ecosystem.

The takeaway: treat every prompt as if it could become public, permanent, and profitable to someone else. That discipline matters beyond ChatGPT, too — if you're connecting agents and third-party skills to your AI stack, vet what they can access. SkillzBench is building free security scanning and benchmarking for exactly that purpose, so you can equip agents with skills that are verified safe before they touch your data.

The Never-Share List: 7 Categories to Keep Out of ChatGPT

The good news: you don't have to guess what's risky. The research is remarkably consistent about which categories of information should never touch consumer ChatGPT. The bad news: most people are already sharing them. LayerX research found 77% of employees paste data into GenAI prompts, averaging around 14 pastes a day — at least three of which contain sensitive material.

Here is your never-share list, synthesized from the evidence:

• Customer and personal data (PII) — LayerX found 40% of files uploaded to GenAI tools contain PII or payment card data, and Zscaler flagged 410 million DLP violations for ChatGPT alone, including attempted Social Security number shares.
• Proprietary source code — Samsung engineers famously leaked confidential semiconductor code this way; 65% of Samsung employees afterward expressed serious security concerns.
• Financial projections and strategic planning documents — including product roadmaps and M&A material.
• Health information — even innocuous queries like low-sugar recipes can get you classified as "health-vulnerable" by inference algorithms.
• Credentials, legal matters, and HR disputes — a 2024 infostealer breach exposed 225,000+ OpenAI credentials on dark web markets.

Why so strict? Stanford HAI researcher Jennifer King warns that sensitive information may be collected and used for training, even if it's in a separate file you uploaded during the conversation. And deletion offers no escape hatch: during the NYT v. OpenAI litigation, a court order forced OpenAI to preserve even deleted consumer chats. Treat anything typed into consumer ChatGPT as potentially permanent.

The workaround is simple and costs you nothing in output quality: anonymize or fictionalize. Replace real customer names with "Customer A," swap actual revenue figures for rounded placeholders, and describe code architecture in prose rather than pasting the repo. ChatGPT reasons about structure and pattern just as well with fictional examples — you get the same insight without the exposure.

The same caution should extend to your wider AI stack, not just the chat window. If you're building agent workflows, the tools and skills you connect matter as much as what you type — published research found roughly a quarter of agent skills carry at least one vulnerability. SkillzBench is building a free directory with security scanning and benchmarking for exactly this reason, so teams can vet what they equip before it ever touches sensitive data.

One final reality check: paying for ChatGPT Plus does not buy corporate-grade privacy — Free and Plus plans train on prompts by default. Anonymization isn't paranoia; on the consumer tier, it's the only reliable protection you have.

Your Defense Playbook: Settings and Habits That Actually Protect You

Knowing what not to share is half the battle. The other half is configuring ChatGPT so your defaults stop working against you — because on consumer plans, the defaults are not on your side. Free and Plus conversations are used for model training unless you actively opt out, and history persists indefinitely until you delete it, according to a consumer privacy comparison that ranked ChatGPT the least private of four major chatbots by default.

Start with the training toggle. In Settings, turn off "Improve the model for everyone" so your prompts stop feeding future model versions. Stanford HAI lead researcher Jennifer King advises users to "whenever possible, affirmatively opt out of having their data used for training," noting that even files you upload during a conversation may be used for training (Stanford HAI).

For genuinely sensitive work, use Temporary Chat. Per OpenAI's own documentation, temporary chats don't appear in your history, don't update Memory, and aren't used for training — though they may be retained up to 30 days for safety purposes.

• Turn off "Improve the model for everyone" in your data controls.
• Use Temporary Chat for anything personal or confidential.
• Avoid thumbs up/down feedback on sensitive chats — it can push the entire conversation into training data even if you opted out.
• Delete Memory entries regularly, since Memory builds a persistent profile across conversations.
• Route business-sensitive work through managed Business or Enterprise workspaces, which OpenAI does not use for training by default.

Watch the feedback trap. Many users don't realize that clicking thumbs up or thumbs down on a chat can send that conversation into training data regardless of your opt-out — a detail buried in OpenAI's data controls documentation. On sensitive threads, skip the ratings entirely.

Also, don't assume your subscription protects you. Only Pro-tier subscribers get automatic training exclusion; Plus subscribers train by default, and paying €20 a month does not buy corporate-grade privacy. For teams, the right move is routing confidential work through ChatGPT Business or Enterprise workspaces, where OpenAI commits to not training on content by default.

Finally, calibrate your expectations about deletion. Deleted chats follow a 30-day deletion window that can be extended "for legal or security reasons," and the NYT v. OpenAI litigation proved a court order can force preservation of even deleted consumer chats (OpenAI). The practical rule: treat anything typed into consumer ChatGPT as potentially permanent.

This discipline extends beyond the chatbot itself. As teams connect agents to third-party skills and tools, the same vetting instinct applies — SkillzBench is building free security scanning and benchmarking to help teams verify which agent skills are safe to plug into their stack. Governance beats prohibition: with 77% of employees already pasting data into GenAI prompts, the goal is safer habits, not abstinence.

For Teams: Govern, Don't Ban — and Vet Your Whole AI Stack

Banning ChatGPT at work feels decisive, but the data says it's a losing strategy. Workplace research found 77% of employees paste data into generative AI prompts, and 82% of those pastes come from personal accounts outside company control — meaning a ban simply pushes usage into the shadows.

The smarter move is governance over prohibition. Employees already average around 14 pastes per day, at least three containing sensitive data, so the question isn't whether AI touches your workflows — it's whether it happens on your terms.

Start with an approved-use policy that names categories employees must never paste into consumer AI tools: customer and personal data, proprietary source code, financial projections, strategic documents, and health information. Then give them a sanctioned alternative, because OpenAI's own documentation confirms that ChatGPT Business and Enterprise workspaces are excluded from model training by default — a protection consumer Free and Plus plans don't offer.

• Define what's off-limits: PII, source code, financials, strategy, health data.
• Route sensitive work through managed, non-training workspaces.
• Train employees on Temporary Chat and training opt-outs for edge cases.
• Audit the full AI stack, not just the chatbot everyone talks about.

That last point matters more than most managers realize. If your teams run AI agents, every third-party skill connected to those agents is another potential doorway into your data. Published research found roughly 26% of agent skills carry at least one vulnerability, yet most teams install them with no scanning or verification at all.

Vetting the whole stack doesn't require a security team. This is exactly the gap SkillzBench is building to close: a free directory of agent skills with version pinning, security scanning, and benchmarking, plus a free MCP connector so your agents only see the skills you've chosen and verified. Every skill gets a trust score from security, quality, license, and maintenance signals — and money never touches scores.

Govern, provide sanctioned tools, and vet everything you connect. Join the SkillzBench waitlist to stay current as the directory and MCP connector roll out.