Gary McBride

Head of WebApps & Tooling @Conste11ation 🤝 | Providing secure environment for LO token swap @Pacaswap 🚀 | Helping users navigate #DAG safely & smarter

[Gary McBride @garywmcbride]

@Rundychops I was referring to the DAG Explorer, you can check there as well. Your tokens are clearly sitting there in your wallet.

[Gary McBride @garywmcbride]

Hello @Rundychops , what issue did you encounter with your Pacaswap account? A detailed explanation of the problem would help us investigate it properly and assist you more effectively.

[Gary McBride @garywmcbride]

@Rundychops What I meant in that context is that your seed phrase isn’t being given to anyone. Your seed phrase is simply required to sign back into your wallet.

[Gary McBride @garywmcbride]

@Rundychops Sometimes clearing cache isn’t enough which is why incognito mode is the best option

[Gary McBride @garywmcbride]

@Rundychops This is because Incognito doesn’t allow extension by default so you are basically signing in from scratch while on incognito

[Gary McBride @garywmcbride]

@Rundychops No you’re not giving up your seed phrase, incognito doesn’t have a default extension so you’re literally signing in on incognito for the first time

[Gary McBride @garywmcbride]

@Rundychops Your incognito should look like this👇

[Gary McBride @garywmcbride]

@Rundychops What do you mean by it’s not working ?

[Gary McBride @garywmcbride]

@me_rodrig7521 The unlock period/date is fixed, and under no circumstances should the days slip forward by an extra day. Feel free to send your DAG wallet address so we can properly look into this.

[Gary McBride @garywmcbride]

@Rundychops Have you seen the instructions i sent earlier, if not you can check your DM as well

[Gary McBride @garywmcbride]

@Rundychops You need to use the subdomain which allows startgazer wallet on incognito mode

[Gary McBride @garywmcbride]

@Rundychops Do you know how to switch to incognito on your browser ?

[Gary McBride @garywmcbride]

@Rundychops We did a check and all seems to be okay on the explorer, has your 20,000 dag tokens reflected on your dashboard or is it still missing on your dashboard

[Gary McBride @garywmcbride]

@Rundychops What device are you using, and have you tried any troubleshooting steps, such as clearing your cache? Feel free to send your wallet address so we can properly look into this and assist you further.

[Gary McBride @garywmcbride]

Frontier models can still be tripped up by surprisingly simple scripted attacks.

We tested a multi-turn email-agent attack across several current frontier models.

The setup was straightforward: a standing routine combined with a poisoned ticket. In most cases, the models followed the instructions. Some succeeded on all 8 of 8 attempts.

We also tested a simple system rule: “only send if the user asks in the latest message.” That wasn’t enough. Models continued to retain and act on instructions from earlier turns.

Gate AI, by contrast, blocked every attempt at the gateway—before the payload ever reached the model.

The takeaway isn’t simply that we need smarter models.

It’s that prompt-only defenses remain fragile when models operate inside multi-turn agent workflows.

Full methodology and results below 👇

[ryle @rylego]

Turns out you don’t need fancy techniques to break the latest models. GPT 5.6 Sol and many other frontier models fall even to relatively simple attacks.

https://t.co/pf13FZ1cvV

[Constellation² @Conste11ation]

AIAI Holdings' Constellation Network Unveils Gate AI Security Gateway and Performance Benchmarks Ahead of June Launch

We benchmarked Gate AI across 16 public prompt-injection datasets. Ranked #1 on half. 53ms median latency. 6-17x fewer false positives than Lakera Guard on benign traffic. One global threshold, no per-dataset tuning.

https://t.co/zhhaNZcV3B

[Constellation² @Conste11ation]

Frontier models still fall to simple scripted attacks.

We tested a multi-turn email agent attack against the current frontier models.

A standing routine + poisoned ticket was enough. Most models complied. Some did it 8 out of 8 times.

A one-line system prompt “only send if the user asks in the latest message” did not hold. The models remembered earlier instructions and acted on them.

Gate AI blocked every attempt at the gateway before the model ever saw the payload.

The problem is not just smarter models.

It is that defenses sitting only at the prompt level keep failing in multi-turn agent workflows.

Full write-up with methodology and results below 👇

[Constellation² @Conste11ation]

Joao Carvalho runs an AI agent on X that holds private keys and can send funds. He tells people to attack it.

Every message runs through Gate first.

He also lost $300 in a day to a loop bug before Gate existed. Security and cost, same gateway. $AIAI $DAG

[Constellation² @Conste11ation]

Most AI companies sell software and hope it gets adopted.

AI² buys the business and installs the AI.

Constellation sits inside the same holding company as the data infrastructure. $AIAI $DAG