Lightning Talks: Eito & John — Data Security & Video Generation

Transcript

This transcript was generated by AI and may contain mistakes, including in names and technical terms. It has not been manually verified. Please refer to the recording for accuracy.

Source: YouTube automatic transcript. Select a timestamp to watch that moment on YouTube.

1. 0:00

   Yeah, I guess uh while the slides are loading um who here h struggled with security when they're building you know agents for other people or customers?

2. 0:10

   Maybe raise of hands.

3. 0:12

   Yeah. So like maybe 10 people around the front. Um okay, there we go. Uh yeah, so I want to talk a little bit about data

4. 0:19

   security um and how data security is the bottleneck for modern AI. Uh why am I qualified to talk about this? I used to work in reinforcement learning research

5. 0:28

   and robotics research. And then uh now I work on sealgate which is all around agent access management. We apply learnings from what I do on the side

6. 0:36

   which is advise frontier labs before they release new products or models and assess the security and then sort of

7. 0:43

   apply a lot of that learning so that it can be democratized to let's say non-lab aligned products right uh which is what to get is about. Um if there's any one

8. 0:52

   thing to take away from this talk uh for agent builders, it's probably this which is you know how you should build uh agents uh that are sort of enterprise

9. 1:01

   ready which is broadly something like the agent should be in the outer harness of the sandbox. Uh assuming your agents

10. 1:07

    uh need to execute code. Uh the agents should execute something like code mode where you know you sort of interact with

11. 1:16

    the external world uh through an MCP gateway. you sandbox the network connections to the outside world so it

12. 1:23

    cannot exfiltrate anything or it cannot uh sort of access things uh without any permission and put an MCP gateway in

13. 1:31

    front uh for any interactions that the agent has with the external world. Maybe some of this is reminiscent of the recent open AI hugging face attacks uh

14. 1:39

    that you might have seen of agents sandbox escape and in my personal opinion this is the best architecture to sort of prevent that uh kind of thing

15. 1:48

    happening. Although obviously it's not a silver bullet and any vulnerabilities in the MCP gateways will um screw this up.

16. 1:55

    Uh I'm going to skip over this. Um yeah, I I'll sort of broadly talk about why MCP gateways are valuable. And you know, it's not just me. If you if you don't

17. 2:03

    want to take it from me, take it from, you know, Harvey, Anthropic, and Sierra who have all built a version of the MCP gateway. And this is sort of becoming

18. 2:09

    the industry uh consensus around how you manage agent access. And a lot of this looks exactly the same. Broadly it

19. 2:17

    basically comes down to okay like it's pretty trivial to vibe code a agent that does some useful stuff but then actually when you wanted to put put it down to

20. 2:26

    production you know you get bottlenecked by a lot of these uh where's the questionnaires where's the CM streaming where's the arback where's the roles

21. 2:34

    right and a lot of this um is not just a technological problem right like the technological problem is a relatively

22. 2:41

    easy and solvable one but it's also a politics problem and I'll sort of talk about that uh within the time I have and you know within MCP gateways actually

23. 2:49

    there's quite a lot sorry within like uh production ready enterprise ready agents you have all this sort of bottom layer that's kind of annoying and very hard to

24. 2:57

    wipe code um that I'll sort of be talking about I'm going to skip over this this is basically just around one of the data security risks which is the

25. 3:05

    lethal trifecta it's been pretty well covered um basic which is basically if your agents can read private data and

26. 3:13

    can like send data to some location you can data leaks. This can happen between outside malicious stackers and also uh

27. 3:21

    internal people like for example let's say reading some uh ELP and then sending it to general on Slack or something like this. Um, yeah, I want to talk uh I'll

28. 3:30

    skip this, but broadly, you know, the sort of data security question is always this sort of trade-off between okay, how much autonomy do you give your agents

29. 3:39

    and how useful are they versus how secure is it? And what we want to do, right, uh, as an industry is basically pursue that parto frontier of how do you

30. 3:47

    have security while having autonomy so that you don't because um, in some sense, right, having a approve every single time is technically secure, but

31. 3:55

    it's actually not, right? when you consider the human fallacies of um clicking approve approve approve. Uh I'm going to skip over this. Um yeah, let's

32. 4:04

    talk a little bit about agent identity which is um I think a pretty hot topic.

33. 4:11

    Uh and this is also another thing that you should encode within your a uh agent gateway. This is how we handle it at segate. But we basically have it to

34. 4:20

    where you know if you have a ooth identity attached right to some of the tools that you call then you attach it

35. 4:27

    to the human identity. But then for some agents right that you would custom configure you would give it sort of least privilege access um into okay this

36. 4:36

    agent in this particular chat session has access to these tools to complete these very specific tasks. And this sort

37. 4:43

    of least privileged um access management is really what you want to aim for when you're designing agents. And then sort of be able to stream all of that design

38. 4:52

    uh sort of agent data access into a single centralized dashboard which is what many of the enterprises will uh essentially demand uh so that they can

39. 5:00

    sort of collect all that logs and make sense of it through agents otherwise via CM or some other system so that they know roughly what the agents doing and

40. 5:09

    they're traceable and trackable and explainable. Um, but then there's a lot of data security details that you got to get right. For example, not uh printing PIIs into the logs, for example, right?

41. 5:20

    And there's actually quite a lot of really annoying things. Um, what the main three reasons I would say around like why you should use an MCP gateway

42. 5:29

    is probably around observability, access control, and security and governance.

43. 5:33

    And then how do you make that easy for uh mostly IT admins and security leaders? But there's also quite a lot of

44. 5:40

    really annoying aspects uh with multiplayer politics where sort of within large enterprises you have different stakeholders that you got to

45. 5:49

    manage and it this is where it gets becomes a non-technical problem but more of a security problem of okay like how do you deal with these different parties

46. 5:57

    and then how do you build technology sort of you know different uh products that resonate for each which is a really boring problem uh but this is why um you

47. 6:07

    know this is a lot of the boring stuff that can move the needle on uh sort of accelerating your timelines for you know agent deployment from say uh you know

48. 6:16

    six months to more like a few weeks right um yeah let's talk a little bit

49. 6:22

    about the um the open AI hugging face incident right which is um it broadly

50. 6:29

    happened uh if you know through the agents within the sandbox being given a impossible task that they cannot solve

51. 6:36

    and then what they did uh was they you know many of these agents had an impossible task and then they found an

52. 6:44

    exploit within Artifactory which is the sort of npm equivalent and then found uh remote code execution ability and then

53. 6:52

    found API keys and then found tokens to basically break out into the internet and then attack hugging face. And um this I think is like a best example of

54. 7:01

    why you should never give agents the ability to access API keys because API keys enable exactly scenarios like this,

55. 7:10

    right? And you really want to build a lot of uh like systems to make sure that all the credentials are hidden away,

56. 7:17

    right? In systems like MCP gateways so that one the uh credentials are never leaked into the logs, right? which can

57. 7:25

    lead to all sorts of issues later and two to make sure that the agents cannot abuse the API keys which are typically overprivileged and then when they are

58. 7:33

    given an impossible task for many of these frontier models you know they're like uh uh you know they are there trained to make you happy right like no

59. 7:41

    matter at all costs even if they hack something you know if they are told to do something they will do it at all costs and um this is the other reason to

60. 7:49

    really use MCP gateways to make sure that the credentials are never writed um yeah I Hope I sort of uh went you know this is a very quick talk but uh hope

61. 7:57

    hopefully I can uh manage to convey why you want MCP gateways around sort of yeah you know to satisfy a lot of the

62. 8:05

    corporate slop politics uh but also to prevent agents from you know using and exploiting and you know be the se uh the

63. 8:13

    pleasure- seeeking missile that the agents are to please you uh and break out into the internet and then obviously the sandbox and network isolation is

64. 8:21

    sort of the the best form factor although I'm not going to claim that any system is perfect. The frontier models are in some sense way smarter than us in

65. 8:29

    these specific areas. Um yeah, that is it from me. Uh yeah, thank you very much for listening. Thank you.

66. 8:37

    And uh yeah, if there are any questions um from prompt to director. So um as we covered uh actually using AI to generate video actually is quite uh cumbersome.

67. 8:49

    Um, so, um, so what, um, uh, so, um, you get beautiful clips

68. 8:57

    initially and you think, "Oh, they're so wonderful." Um, but you actually start to do context such as character drift.

69. 9:03

    Um, you sometimes you get unwanted con uh content. You get camera inconsistencies uh, and uh, reference mismatch, a whole

70. 9:12

    heap of things that are very difficult to to overcome. So, so what do we do? Um

71. 9:19

    so we have uh a agent uh orchestrate we've got an orchestration agent uh

72. 9:28

    called um a director agent and within that we have sub agents that actually um

73. 9:34

    specifically do uh specific tasks. So uh we have a camera agent, a lighting agent, continuity agent, performance

74. 9:42

    agent and uh um editor agent and sound agent. And all these are orchestrated by the director agent and and what we've

75. 9:51

    actually done is we've actually started working with Hollywood um Hollywood producers um to actually go through how

76. 9:58

    they think rather than what we think is the best solution. And the Hollywood producers actually introduce things that

77. 10:07

    we just wouldn't have thought about. um uh things like uh h how how the cameras move, how the uh the the actual lenses

78. 10:16

    um and and so what we've been doing is going through all these very very specific industry things that literally have been um in the industry for over a

79. 10:25

    hundred years and we're going through them piece by piece and then we're we're building agents that actually map on to all these corresponding parameters.

80. 10:36

    Um um so um it it's more than just um 8 seconds of uh of of clip. Uh and uh

81. 10:46

    usually most uh video generation uh relies on about 8 seconds. After that it

82. 10:53

    starts to get very inconsistent. So um and but just as we've been with other

83. 10:59

    talks we've been talking about uh the context. Uh context is everything.

84. 11:08

    Um so how how it works one one hex okay creative intent you put the creative intent in the director agent

85. 11:16

    we've already discussed uh we have a tool orchestration um it calls up all the the different tools that are required to actually physically do the

86. 11:25

    the uh generation and then there's a generation runtime um and then there's a generated uh video and we use all these

87. 11:34

    uh below you can see all the all the um hardware, all the different uh code bases that we actually uh draw upon. Um

88. 11:43

    um and um yeah, I think that probably gets us there. All right. Thank you so much.