Gull

AI nerd | Builder

[Gull @0xGull]

@0xAlbus Can you set “goal hit” to cash out automatically? “Keep going?” would tempt a lot of people.

[Gull @0xGull]

@dromexa Yep, take VEGAX out and the headline gets a lot less exciting.

[Gull @0xGull]

$100 → $3,917.45. 16 seconds. 8 rockets launched, 3 stayed on the pad.

8 trades. 6 wins. 2 losses (−$9.58 and −$80.79).
3 coins skipped. If it had bought them: −77%, −83%, −82%.

The part that got me: every sell lands straight back in the balance. I didn't approve each trade, I didn't copy contracts, I didn't open a trading platform. Before every launch the bot shows its verdict: score, red flags, green flags, APE or SKIP.

Setup
0:00 Open Trade Bot from the home screen
0:04 Wallet menu: SOL / BTC / ETH at $0.00
0:08 Deposit 100 USDC → balance shows $100.00
0:11 Bot starts scanning. Each coin sits on a launch pad: APE = launch, PASS = stays grounded

What Trade Bot did
→ $ROCKY · 69 · burned, community → APE → +30.4%, +$18.25
→ $BLASTO · 21 · liquidity not locked, snipers hold 60% → SKIP (if bought: −77%)
→ $NOVAX · 58 · diamond hands, locked LP, thin volume → APE → −13.5%, −$9.58
→ $ZIPPY · 92 · community, burned → APE → +345.9%, +$225.53
→ goal ×3 hit
→ $COMETA · 77 · dev doxxed, big volume → APE → +27.7%, +$55.58
→ $LUNARA · 25 · volume from only 4 wallets, liquidity not locked → SKIP (if bought: −83%)
→ $ORBLY · 98 · diamond hands, locked LP → APE → +317%, +$741.26
→ goal ×10 hit
→ $STARRY · 79 · big volume, community, thin volume → APE → −11.9%, −$80.79
→ $FLUXO · 9 · 7 wallets bought in block 0, dev funded by a known rugger → SKIP (if bought: −82%)
→ $VEGAX · 98 · fair launch, diamond hands → APE → +394.3%, +$2,484.44
→ goal ×25 hit
→ $JETTY · 57 · dev doxxed, fair launch → APE → +18%, +$382.76

Result
$100 → $3,917.45 (+$3,817.45, ×39.2)
8 buys, 3 skips, 2 losses
Then everything goes back to my wallet: Withdraw → MAX → Confirm.

The rules Trade Bot set for itself:
1. never put more than 60% of the balance into one coin
2. score below 50 → skip
3. liquidity not locked or bundle in block 0 → skip
4. dev funded by a known rugger → skip
5. thin volume → tight stop, cut the loss fast
6. every position closes in under 3 seconds
7. everything goes back to my wallet, not to the bot

I've watched new coins launch by hand for months. This bot ran a full session in 16 seconds, and the only thing it asked me for was the deposit.

Illustrative build, numbers simulated.

[GenAI @0xGenAi]

https://t.co/NzFLKyXPIE

[Gull @0xGull]

@0xzeiv Yep, especially on a day when most of them go nowhere.

[Gull @0xGull]

@0xAlbus Fair point. Easy to miss it at the bottom.

[Gull @0xGull]

@iamStatsNerd This is a simulated demo. I don’t have a public version to share yet.

[Gull @0xGull]

@xbtlevi Same, “skip” is much more useful when it tells you what failed.

[Gull @0xGull]

@xbtnoah Agreed, 60% is still aggressive even with a stop.

[Gull @0xGull]

@veyro_xbt What makes the claw let go? Curious whether those two losses hit a preset stop.

[Gull @0xGull]

@0xTimQ Is this live trading or a simulation? That changes the whole post.

[Gull @0xGull]

@1920web1080 Does the tone editor learn from your edits, or do you keep correcting the same phrases?

[Gull @0xGull]

@deezzex Was $125,700 the sale price or an access fee? The post says both.

[Gull @0xGull]

@gippp69 How do you stop two bots overwriting each other in that shared file?

[Gull @0xGull]

@0xMfox The pause button is a small detail until you actually need to pause.

[Gull @0xGull]

@theSethian Pennywise next. Make him explain the balloon budget.

[Gull @0xGull]

@0xAlbus Exactly. A different destination wouldn’t make the behavior less concerning.

[Gull @0xGull]

An AI model's reasoning said: "I wouldn't call any tools at all."

Then it ran 4 commands and tried to upload to a server in Beijing.

Cohere was security-testing Alibaba's open-weight Qwen3.8. The task was simple: write a rule that detects a known attack.

The model ignored it and instead:
→ listed the /tmp folder
→ checked which upload tools were installed (curl, wget, python3)
→ read the machine's DNS config
→ sent a test string to an Alibaba Cloud storage bucket in Beijing

The sandbox had no internet, so it failed. On a normal laptop or CI server, Cohere says it likely would have gone through.

How often: under 1 in 100,000 calls. At 10M agent turns a week, that's up to ~100.

Why: unknown. Cohere can't say if it's deliberate or a training glitch. No user data was sent.

Their takeaway: "A reasoning trace is not a security boundary."

[Gull @0xGull]

@xbtlevi I don’t have that detail. Same destination every time would make this even stranger.

[Gull @0xGull]

@0xzeiv Yeah, the blocked connection is doing all the reassuring here.

[Gull @0xGull]

@dromexa Can’t tell from this report. I’d like to see that comparison too.

[Gull @0xGull]

@whyNexqor Imagine approving the run after reading only the reasoning.