1
00:00:04,909 --> 00:00:08,539
US orders Anthropic to block foreign access to top models.

2
00:00:08,869 --> 00:00:11,609
Compliance teams at every frontier lab are already

3
00:00:11,609 --> 00:00:14,250
modeling the knock-on effects of restricted access.

4
00:00:14,550 --> 00:00:18,829
OpenAI must decide whether to mirror the blocks or absorb the lost international

5
00:00:18,829 --> 00:00:22,629
traffic that currently accounts for nearly a third of its API volume.

6
00:00:23,340 --> 00:00:26,600
European research groups will start spinning up US based proxies,

7
00:00:26,819 --> 00:00:29,559
pushing average query costs up 15%

8
00:00:29,700 --> 00:00:32,168
while introducing new data residency headaches.

9
00:00:32,459 --> 00:00:36,580
The split in model access hardens faster than any export rule intended.

10
00:00:37,279 --> 00:00:41,919
Salesforce acquires FinAI agent platform for $3.6 billion.

11
00:00:42,240 --> 00:00:45,279
ServiceNow faces an immediate choice on whether to release its

12
00:00:45,279 --> 00:00:49,349
own resolution benchmarks or concede ground in upcoming RFPs.

13
00:00:49,639 --> 00:00:53,959
The 76% figure sets a new baseline that forces every help desk vendor.

14
00:00:54,014 --> 00:00:57,884
To reprice human escalation paths within existing contracts,

15
00:00:58,115 --> 00:01:02,474
mid-market buyers will demand similar autonomy clauses in the next two quarters,

16
00:01:02,645 --> 00:01:05,535
compressing implementation cycles across the category.

17
00:01:05,894 --> 00:01:08,535
Legacy vendors without native agents start shopping

18
00:01:08,535 --> 00:01:10,915
for acquisitions before their next earnings call.

19
00:01:11,739 --> 00:01:16,290
New Corp raises $66 million for AI agent employee identities.

20
00:01:16,610 --> 00:01:21,050
Enterprise security reviews now treat AI agents as new hire equivalents,

21
00:01:21,339 --> 00:01:24,330
complete with access revocation timelines and audit trails.

22
00:01:24,610 --> 00:01:26,220
This raises the bar for any team

23
00:01:26,220 --> 00:01:29,959
shipping autonomous workflows without pre-approved identity controls.

24
00:01:30,260 --> 00:01:33,680
Vendors lacking similar tooling will see deployment cycles extend

25
00:01:33,680 --> 00:01:37,809
by months as procurement adds dedicated agent governance checkpoints.

26
00:01:38,099 --> 00:01:41,809
The $300 million valuation already prices in mandatory

27
00:01:41,809 --> 00:01:45,139
adoption across regulated industries within 18 months,

28
00:01:45,300 --> 00:01:48,540
shifting budgets from model training to compliance infrastructure.

29
00:01:48,940 --> 00:01:52,330
Anthropic proposes advanced AI governance frameworks.

30
00:01:52,500 --> 00:01:56,900
OpenAI will have to match the transparency and evaluation standards or

31
00:01:56,900 --> 00:02:00,459
explain to policymakers why its own models skip comparable scrutiny.

32
00:02:00,819 --> 00:02:03,139
Worker support provisions in the framework create

33
00:02:03,139 --> 00:02:05,169
new cost baselines that smaller labs.

34
00:02:05,305 --> 00:02:07,735
Cannot absorb without raising API rates.

35
00:02:08,024 --> 00:02:11,925
The documents also quietly favor labs with existing safety teams,

36
00:02:12,024 --> 00:02:13,904
giving Anthropic an edge in any

37
00:02:13,904 --> 00:02:16,524
government procurement that adopts these criteria.

38
00:02:16,744 --> 00:02:19,785
Expect similar releases from competitors inside 90

39
00:02:19,785 --> 00:02:21,554
days to stay in the conversation.

40
00:02:22,065 --> 00:02:25,505
Fable 5 120K system prompt leaked on GitHub.

41
00:02:25,899 --> 00:02:29,000
Those safety classifiers were designed to evolve in secret.

42
00:02:29,220 --> 00:02:31,539
Yet the complete specification now sits in a

43
00:02:31,539 --> 00:02:34,100
public repo for anyone to benchmark against.

44
00:02:34,460 --> 00:02:36,770
Fine tuning teams will score their models on the

45
00:02:36,770 --> 00:02:39,970
exact refusal triggers instead of guessing at categories.

46
00:02:40,279 --> 00:02:42,960
Smaller labs copy the structure directly now.

47
00:02:43,179 --> 00:02:45,899
Expect open source projects to match Fable 5

48
00:02:45,899 --> 00:02:48,619
consistency on edge cases within 9 months,

49
00:02:48,860 --> 00:02:51,020
undercutting any edge from keeping the rules

50
00:02:51,020 --> 00:02:54,139
internal and accelerating their next rewrite cycle.

51
00:02:54,750 --> 00:02:58,190
Multi-agent pack hunt jailbreak bypasses Fable 5.

52
00:02:58,550 --> 00:03:01,830
Coordinated agent swarms expose the limits of single prompt

53
00:03:01,830 --> 00:03:05,020
classifiers faster than any solo attack ever could.

54
00:03:05,350 --> 00:03:07,509
Defenses built for single sessions break

55
00:03:07,509 --> 00:03:09,970
when multiple models trade partial answers.

56
00:03:10,190 --> 00:03:12,110
The practical response is heavier per

57
00:03:12,110 --> 00:03:15,050
session logging and cross-agent correlation checks

58
00:03:15,149 --> 00:03:18,830
that anthropic must now run at scale on every automated workflow.

59
00:03:19,220 --> 00:03:21,720
Those checks raise inference cost per query,

60
00:03:21,940 --> 00:03:24,580
so expect the programmatic credit price to climb within

61
00:03:24,580 --> 00:03:27,460
two quarters just to cover the added safety overhead.

62
00:03:28,100 --> 00:03:31,320
Cloud billing shifts to programmatic credits June 15th.

63
00:03:31,610 --> 00:03:34,449
Separate credit pools let anthropic charge agent runs

64
00:03:34,449 --> 00:03:38,059
at full API rates without touching consumer subscriptions.

65
00:03:38,289 --> 00:03:40,919
The move isolates high volume automated traffic,

66
00:03:41,089 --> 00:03:44,559
so it cannot keep pulling down per token margins on the chat side.

67
00:03:44,850 --> 00:03:48,880
Smaller teams running continuous agents now see costs rise separately.

68
00:03:49,169 --> 00:03:49,899
Competitors face

69
00:03:50,026 --> 00:03:51,716
The same choice inside a year.

70
00:03:51,945 --> 00:03:53,826
Keep unified billing and watch agent

71
00:03:53,826 --> 00:03:56,126
workloads erode their highest margin product,

72
00:03:56,425 --> 00:03:59,526
or copy the split and explain two ledgers to users.

73
00:04:00,026 --> 00:04:03,406
Cyber experts urge reversal of anthropic restrictions.

74
00:04:03,626 --> 00:04:05,945
Reversing the curbs would let foreign researchers

75
00:04:05,945 --> 00:04:08,526
contribute to US aligned safety benchmarks again.

76
00:04:08,985 --> 00:04:11,955
The administration's decision now determines whether Allied

77
00:04:12,042 --> 00:04:14,901
Labs stay inside the current evaluation ecosystem.

78
00:04:15,242 --> 00:04:17,221
Keeping them in place past next spring,

79
00:04:17,562 --> 00:04:20,182
instead hands competing labs in Europe and Asia

80
00:04:20,321 --> 00:04:23,002
an uncontested period to build and validate

81
00:04:23,002 --> 00:04:25,782
sovereign alternatives without any US input.

82
00:04:26,041 --> 00:04:28,782
Those alternatives then set their own access rules,

83
00:04:28,872 --> 00:04:31,201
reducing US influence over the models that

84
00:04:31,201 --> 00:04:34,191
end up deployed in critical infrastructure worldwide.

85
00:04:34,601 --> 00:04:37,742
Anthropic meets White House on model security concerns.

86
00:04:37,962 --> 00:04:40,641
Export controls are no longer the main constraint.

87
00:04:40,920 --> 00:04:43,399
The talks center on whether Anthropic must now

88
00:04:43,399 --> 00:04:46,739
host separate model instances for each sanctioned region,

89
00:04:46,940 --> 00:04:50,559
a step that would double its infrastructure costs within 18 months.

90
00:04:50,859 --> 00:04:53,559
Smaller developers migrate first because they lack the

91
00:04:53,559 --> 00:04:56,140
legal teams to navigate per country rules.

92
00:04:56,399 --> 00:05:00,160
That shift hands share to providers accepting fragmented deployments.

93
00:05:00,480 --> 00:05:04,279
Anthropic's single stack model cannot survive the new compliance layer.

94
00:05:04,820 --> 00:05:08,940
MetaCTO calls AI reorg atrocious amid unrest.

95
00:05:09,179 --> 00:05:12,519
Andrew Bosworth's frank assessment masks a deeper problem.

96
00:05:12,700 --> 00:05:16,339
The 6500 person unit has already lost its two

97
00:05:16,339 --> 00:05:19,839
lead architects to startups building competing agent frameworks.

98
00:05:20,019 --> 00:05:22,380
Retention offers failed because the reorg

99
00:05:22,380 --> 00:05:24,959
removed direct ownership over model decisions.

100
00:05:25,179 --> 00:05:28,079
Expect Google and OpenAI to accelerate their own

101
00:05:28,079 --> 00:05:31,179
internal agent teams while Mehta rebuilds basic trust.

102
00:05:31,410 --> 00:05:35,230
The coding focus now looks like a distraction from the real capability gap.

103
00:05:36,040 --> 00:05:39,829
KPMG pulls agentic AI report over hallucinations.

104
00:05:40,119 --> 00:05:42,959
Client pushback on fabricated case studies has already

105
00:05:42,959 --> 00:05:46,019
triggered internal audits at every major consultancy.

106
00:05:46,359 --> 00:05:50,760
KPMG's move signals that future agentic AI publications will require

107
00:05:50,760 --> 00:05:54,470
third party verification of every usage statistic before release.

108
00:05:54,799 --> 00:05:58,790
That requirement alone will delay industry benchmarks by at least 2 quarters

109
00:05:58,959 --> 00:06:02,160
and give smaller research outfits an opening to publish first.

110
00:06:02,459 --> 00:06:05,230
The credibility hit extends beyond one document.

111
00:06:05,549 --> 00:06:08,559
AI agents surpass humans in web traffic share.

112
00:06:08,899 --> 00:06:11,549
Cloudflare's traffic split means recommendation

113
00:06:11,549 --> 00:06:13,230
engines must now price impressions

114
00:06:13,230 --> 00:06:16,459
based on agent negotiation patterns instead of dwell time.

115
00:06:16,750 --> 00:06:20,510
Retailers lose the ability to AB test human responses when

116
00:06:20,510 --> 00:06:24,059
over half their visitors arrive without cookies or intent signals.

117
00:06:24,429 --> 00:06:27,109
Watch Amazon's margin on sponsored products once

118
00:06:27,109 --> 00:06:30,040
agents start routing around paid placements entirely.

119
00:06:30,369 --> 00:06:32,989
The shift already forces ad networks to expose

120
00:06:32,989 --> 00:06:35,809
bidding APIs that agents can call directly.