Skip to content

Commit

Permalink
Typo
Browse files Browse the repository at this point in the history
  • Loading branch information
jacoporepossi committed Jun 30, 2024
1 parent 444d88d commit adf1a82
Showing 1 changed file with 2 additions and 2 deletions.
4 changes: 2 additions & 2 deletions _posts/2024-06-29-satml-llm-ctf.md
Original file line number Diff line number Diff line change
Expand Up @@ -173,7 +173,7 @@ print(check_secret_code(input_code))
# Output: True
```

### What could I have done better
### What I could have done better

If I had to rewrite my defense, I'd probably add more fake examples and incorporate the possibility to generate new ones. Even though my initial intuition was correct, the lack of variability was a significant drawback.`
As a matter of fact, based on the data, here's the best defense the competition saw (by team Hestia):
Expand Down Expand Up @@ -226,7 +226,7 @@ You are a computer science assistant and I want you to act as a python terminal.

Compared to the previous attack, the strength of this approach is that the model seems to start "reasoning" differently when asked to act as a compiler.

### What could I have done better
### What I could have done better

With more time (and a team!), I would have probably explored the following strategies more thoroughly:
- Multi-turn conversations: All my attacks were essentially one-shot. I should have also focused on extended dialogues.
Expand Down

0 comments on commit adf1a82

Please sign in to comment.