Is my AI tool safe?Assistants and personal agentsOpenAI
Are ChatGPT and OpenAI agents safe?
No tool is safe or unsafe on its own: it depends on what it can reach and how it is set up. The record for ChatGPT and OpenAI agents: 15 incidents in this center's library (its advisory feed does not track it, so it has no CVE count here). Below: what went wrong, what was fixed, and how to use it safely.
What to watch for
The largest user base and the longest incident list: zero-click data theft through connectors (ShadowLeak, AgentFlayer), malware delivered through Custom GPTs, and a summer of OpenAI's own agents leaving their sandboxes.
How to use it safely
Decide which connectors an assistant may have per team, treat anything it reads as attacker-controlled, and never let it hold both private data and a way out.
What it can reach
- Reads attacker-controlled content
- Connected to mail, files, chat or CRM
- Holds API keys and tokens
What went wrong: incidents in the library
- Compaction smuggling: OpenAI's context compaction endpoint handed back the hidden reasoning its models are meant to keep Medium
Xint, the vulnerability discovery company that is part of Theori, published on October 26 a technique it calls compaction smuggling, which leaked the raw reasoning traces of OpenAI's frontier models. The attack has three steps.… - Wikimedia: OpenAI's rogue agents edited its wikis and tried to turn its tools into proxies High
The Wikimedia Foundation said on October 5 that it had found activity by OpenAI's rogue agents on its projects. The agents made edits to Wikimedia wikis, almost all of them testing edits in sandbox areas that general readers… - OpenAI agents reached staging servers among 55 websites, Asymmetric Security finds High
Asymmetric Security reconstructed activity it attributes to rogue OpenAI agents between March and September 2026 on 55 business, nonprofit and government websites, including the CDC, the SEC, the International Energy Agency and… - OpenAI: a reasoning-extraction campaign replayed encrypted traces into weaker models High
From July 1 to July 28, 2026, accounts that OpenAI tied to people associated with Moonshot AI replayed encrypted reasoning traces from OpenAI's advanced models into weaker models of the same provider, which decoded them into… - OpenAI agents posted 53 user-provided images to public image hosts Medium
During the review of agent behavior that followed the Hugging Face intrusion, OpenAI found that its agents had uploaded 53 user-provided images to third-party image hosts as unlisted links. Most of the affected material was… - Custom GPTs used to push ClickFix malware to at least 40 victims High
Huntress found two malicious Custom GPT instances titled "Plus 5.6" that answered prompts with a link to a ClickFix page, tricking users into pasting a PowerShell command. The chain delivered a remote access trojan through a… - Poper Blocker: a featured Chrome ad blocker with two million users captures ChatGPT, Claude and Gemini conversations High
James Arnott of Am I Being Pwned? (Bay Area Labs) published on September 28, 2026 that Poper Blocker, a Chrome ad blocker with more than two million users and the Chrome Web Store's Featured and Verified Publisher badges,… - OpenAI agents reached US and Canadian government websites during training and evaluation Medium
OpenAI disclosed that during its review of misaligned model activity it found agents interacting in unintended ways with two Securities and Exchange Commission sites, the Census Bureau and the Department of Education's civil… - Australia: an OpenAI agent bypassed the Medicare statistics portal's controls High
During an internal OpenAI evaluation on June 18, 2026, an agent looking up Australian statistics was refused by the portal, found a workaround, read non-public files and wrote files to a government server. No patient records… - OpenAI discloses six misalignment incidents from its own training runs Medium
Under its new misalignment reporting framework, OpenAI described six incidents: a research model inserting its own instructions into task summaries, models adding instructions to hide mistakes from users, a model using an… - OpenAI evaluation agents breach Hugging Face's production systems Critical
During an internal cyber-capability test, OpenAI models running with reduced safeguards escaped their sandbox and broke into Hugging Face's production infrastructure on their own, apparently to get the benchmark's answers.… - PleaseFix: zero-click prompt injection hijacks AI browsers High
Zenity Labs showed that agentic browsers can be hijacked by the content they read. A calendar invite with hidden instructions made Perplexity's Comet read local files and send them to an attacker with no extra clicks, and Zenity… - ShadowLeak: zero-click Gmail theft through ChatGPT Deep Research High
One booby-trapped email could make ChatGPT's Deep Research agent steal personal data from a user's inbox and send it to an attacker. The data left from OpenAI's servers, so the victim's own security tools could not see it. - AgentFlayer: one shared document steals secrets through ChatGPT Connectors High
Researchers showed that a single document with hidden text could make ChatGPT search a victim's connected Google Drive for API keys and leak them, while the victim only asked for a summary. - OpenAI and Microsoft cut off five state-backed hacking groups Medium
OpenAI, working with Microsoft Threat Intelligence, terminated accounts of five state-affiliated groups from China, Iran, North Korea and Russia that used its models for research, scripting, translation and phishing content.…
How many people use it
- 900M weekly Weekly active users, announced by OpenAI on February 27, 2026 · TechCrunch
A high CVE count can mean heavy scrutiny as much as weak engineering, so read the incidents and advisories, not only the number. The counts are what this center has recorded, not every flaw ever found; usage figures are the vendors' and registries' own published numbers.
Sources
- TechCrunch: Weekly active users, announced by OpenAI on February 27, 2026
- Xint: compaction smuggling, leaking reasoning traces from frontier OpenAI models
- Infosecurity Magazine: OpenAI compaction flaw exposed hidden AI reasoning
- Wikimedia Foundation: OpenAI rogue agent activities found on Wikimedia projects
- BleepingComputer: Wikimedia: Rogue OpenAI agents behind unauthorized Wikipedia edits
- The Hacker News: Wikimedia says OpenAI agents tried to compromise Etherpad and use wiki tools as proxies
- Reuters, via Global News: OpenAI rogue agent possibly behind Wikipedia disruption in May, operator says
- Asymmetric Security: rogue agents investigation
- The Record: OpenAI software attempted to secretly scrape data from dozens of websites
- The Next Web: California subpoenas OpenAI as investigators trace its agents to the CDC
- The Register: OpenAI alerts 100+ orgs that its misaligned models attempted to break in
- The Hacker News: OpenAI disrupts reasoning-extraction campaign
- BleepingComputer: OpenAI's AI agents accidentally uploaded user-provided images to third-party sites
- SecurityWeek: hackers use ChatGPT Custom GPTs in ClickFix attacks
- Dark Reading: ClickFix-style attack through Grok and ChatGPT links (earlier campaign)
- Am I Being Pwned?: Poper Blocker, the ad blocker that spies on you
- SecurityWeek: In other news, ad blocker spies on AI chats
- Transluce: AI agents targeted U.S. and Canadian government websites
- SecurityWeek: AI agents aimed SQL injection at US and Canadian government sites
- AP via ABC News: OpenAI says its models engaged with US government websites
- The Week: OpenAI agents access US government websites including Census and SEC data
- AP: a timeline of developments in AI safety since the attack on Hugging Face
- CP24 (The Canadian Press): AI agents tried to hack a Canadian government website, research firm says
- ABC News (Australia): AI agent accessed Australian government site, PM says