Of 5,132 X posts about bots and AI agents, 86 were first-person losses
We searched X for posts about trading bots and AI agents that touch money, from January 2025 to October 2026, and labelled every relevant post by hand. The losses people described came from bots that hold their keys, leaked exchange API keys and scam bots. Almost none were an AI going rogue. The labelled data is open: CC BY 4.0.
5,132posts collected
1,173relevant, labelled
86first-person losses
3from a runaway bot or agent
Download CSVJSONLCollections CSVMethod and themes (JSON)
What we asked
Which problems with bots and AI agents that touch money do people actually post about on X? How often, who posts, and do they ask for a fix? We started from ten guesses (themes) and searched for each one the same way, so the themes can be compared.
- T1 Runaway or unsafe bot or agent actions: wrong orders, no limits, no approval, needing a kill switch or caps.
- T2 Exchange API keys leaked or stolen: keys stolen, pushed to GitHub, or taken in a bot platform breach such as 3Commas.
- T3 Bot custody: Telegram and sniper bots that hold private keys, bot hacks, "the bot drained my wallet", malware posing as a bot.
- T4 Prompt injection or manipulation of crypto AI agents: tricked into sending funds, memory injection.
- T5 Trust and selection: is this bot legit, scam bots, which bot to trust, reviews.
- T6 Fake or unverifiable performance: fake PnL, signal-group scams, backtests that don't hold, wanting verified track records.
- T7 Onchain strategy privacy: wallet tracking, copy-traders following you, front-running or MEV against your trades or bot.
- T8 Data privacy with AI: sharing a portfolio, trades, keys or brokerage access with ChatGPT, Claude, Grok or other LLM tools.
- T9 Agent payments and spending control: agents with cards or crypto overspending, budgets, x402, approval to pay.
- T10 Baseline: does AI trading actually make money (profitability, "doesn't work").
How we collected
- X search in a signed-in browser, read only. We never liked, followed, replied or posted.
- 4 searches per theme, 40 in all. The exact search strings are in the meta file.
- Each search ran 5 times: X's Top tab (all time, no date filter), and the Latest tab in four windows: Jan to Jun 2025, Jul to Dec 2025, Jan to Jun 2026 and Jul 1 to Oct 24, 2026. The dated windows add min_faves:2, so a post needs at least 2 likes.
- For each theme we also opened the 4 most-replied relevant posts (not promo) and collected their replies: 41 thread collections.
- 241 collections returned 5,132 posts. A post counts again if a second search found it.
- A post was kept only if it was really about the theme's problem in trading, crypto, agents or bots. Off-topic matches, general market chatter and unrelated hacks were dropped. 1,173 were kept (23%). One more post was ours and is left out.
How we labelled each post
| Label | Value | Meaning |
|---|---|---|
| type | pain | First-person loss, frustration or fear about the problem |
ask | Asks how to solve it, for a tool or for advice, or says someone should build this or they would pay | |
incident | News or report of a specific hack, exploit, leak or loss affecting others | |
discussion | Opinion, analysis, warning or education | |
promo | A product or account marketing itself as the fix, shilling, giveaways | |
| author_category | retail | Individual trader or user |
builder | Developer, founder or researcher speaking as a builder | |
vendor | Project or company account, or its staff, promoting | |
media | News, alert or research account | |
| intensity | 1 | Mild mention |
2 | Clear problem | |
3 | Severe (money lost, urgent) | |
| seeks_solution | 1 | Asks or wishes for a fix, tool, check or safeguard (0 otherwise) |
| note | What the post says, in our own words (15 words or fewer) |
The 86 first-person losses
- 55 Telegram and sniper bot custody. A wallet drained after its key went into a Telegram or sniper bot, including fake and malware bots posing as real ones.
- 19 exchange API keys. Keys stolen in a bot platform breach, most often the 3Commas API key leak, or pushed to GitHub by mistake.
- 9 scam bots and fake platforms. Money lost to a scam bot, a fake "AI trading" platform or a copy-trading scam.
- 3 runaway bot or agent actions. One trading-bot glitch, and two people whose agents ran in loops.
- 0 from prompt injection. 134 posts talked about prompt injection of crypto agents, but none described a loss of the poster's own.
73 of the 86 were rated severe (money lost or urgent) and 84 came from individual traders. Most of the money people say they lost went through what they connected: a bot holding their key, an exchange key, or a fake bot. That is what TAS Check looks at before you connect.
Results by theme
| Theme | Relevant | Non-promo | Authors | First-person losses | Pain + ask | Solution-seekers | Promo | Demand score | Precision |
|---|---|---|---|---|---|---|---|---|---|
| T3 Bot custody & Telegram bot drains | 134 | 104 | 94 | 55 | 56 | 12 | 30 | 560.6 | 27% |
| T2 Exchange API key leaks/theft | 143 | 127 | 110 | 19 | 25 | 7 | 16 | 452.7 | 29% |
| T6 Fake/unverifiable performance | 219 | 122 | 114 | 5 | 6 | 1 | 97 | 332.1 | 33% |
| T5 Trust: scam bots, which bot is legit | 143 | 103 | 92 | 9 | 11 | 3 | 40 | 313.2 | 25% |
| T4 Prompt injection of crypto agents | 134 | 103 | 98 | 0 | 3 | 16 | 31 | 235.0 | 28% |
| T10 Does AI trading make money | 130 | 57 | 54 | 7 | 7 | 0 | 73 | 146.0 | 25% |
| T9 Agent payments / spending control | 94 | 65 | 64 | 2 | 4 | 4 | 29 | 145.4 | 19% |
| T1 Runaway/unsafe agent or bot actions | 87 | 71 | 70 | 3 | 9 | 10 | 16 | 115.5 | 19% |
| T7 Onchain strategy privacy / front-running | 56 | 16 | 15 | 2 | 2 | 0 | 40 | 45.1 | 12% |
| T8 Data privacy with AI tools | 33 | 13 | 12 | 0 | 1 | 1 | 20 | 28.3 | 7% |
Limits
- X search returns a sample, not every post, so these are not totals.
- min_faves:2 on the dated windows drops posts with fewer than 2 likes.
- The Top tab is X's own ranking, and we can't see how it picks.
- One person labelled every post. Pain also covers frustration or fear about a loss, not only money lost; the intensity label shows how serious each one is.
- Thread replies can be inflated by engagement campaigns. Where we spotted one, its posts were labelled promo or dropped as off-topic.
- The themes are comparable because each got the same searches and windows, not because any of them is complete.
Fields
| Field | Meaning |
|---|---|
post_id | X post ID |
url | Link to the post: https://x.com/<handle>/status/<post_id> |
date | Date posted (UTC) |
theme, theme_name | T1 to T10, see the themes above |
query_id | Which of the theme's 4 searches found it (queries are in the meta file); thread replies keep the query of their thread |
window | TOP, W1 to W4, or thread |
thread_root_id | For thread replies: the post whose replies we collected |
type | pain, ask, incident, discussion or promo |
author_category | retail, builder, vendor or media |
seeks_solution | 1 if the post asks or wishes for a fix, tool, check or safeguard |
intensity | 1 mild, 2 clear problem, 3 severe (money lost, urgent) |
note | What the post says, in our own words (15 words or fewer) |
replies, reposts, likes, views | Engagement as shown on X on Oct 24, 2026 (views 0 = not shown) |
is_reply | 1 if the post is a reply |
Use the data
Free to use under CC BY 4.0: credit "TradingAgentStack (tradingagentstack.com)" with a link to this page. To read a post, open its link on X; we don't republish what people wrote. Found a label you think is wrong? Tell us with the post ID.
More research and data: broker MCP order gating, the incident tracker and the open dataset of AI trading tools. Not financial advice.