文字起こし 🔗
WEBVTT
00:00:03.470 --> 00:00:07.509
In 2022, the median time between initial access
00:00:07.509 --> 00:00:11.609
and secondary thread handoff was 8 hours. At
00:00:11.609 --> 00:00:15.349
RSA conference 26, Google's Mandiant team put
00:00:15.349 --> 00:00:19.390
the new number on stage, 22 seconds. Some breakouts
00:00:19.390 --> 00:00:24.370
now happen in 27 seconds flat. That single statistic
00:00:24.370 --> 00:00:27.449
explains everything that got announced in San
00:00:27.449 --> 00:00:29.690
Francisco this year. This is Tech Updates, let's
00:00:29.690 --> 00:00:34.460
get into it. RSA Conference 2026 ran from March
00:00:34.460 --> 00:00:39.579
23rd to 26th at Moscone. 44 ,000 attendees, 600
00:00:39.579 --> 00:00:44.399
exhibitors, 570 plus sessions, and the team was
00:00:44.399 --> 00:00:47.539
the power of community, but the actual team on
00:00:47.539 --> 00:00:50.820
the show floor was agentic AI, not a buzzword
00:00:50.820 --> 00:00:53.359
as an attack surface and a control plane problem.
00:00:55.049 --> 00:00:57.750
Cisco's Jiddu Patel framed it on the main stage
00:00:57.750 --> 00:01:00.850
with a line word repeating. With chatbots you
00:01:00.850 --> 00:01:03.829
worry about getting the wrong answer, with agents
00:01:03.829 --> 00:01:07.769
you worry about taking the wrong action. That's
00:01:07.769 --> 00:01:10.349
the whole shift. We're moving from LLMs that
00:01:10.349 --> 00:01:13.870
talk to LLMs that do. They call tools, they hit
00:01:13.870 --> 00:01:16.349
APIs, they move money, they touch production,
00:01:16.909 --> 00:01:19.590
and every one of them is non -human identity
00:01:19.590 --> 00:01:23.629
with credentials. A blast radius and zero intuition.
00:01:24.120 --> 00:01:27.459
The Innovation Sandbox winner reinforced this.
00:01:27.760 --> 00:01:31.180
Georgia AI with most innovative startup for an
00:01:31.180 --> 00:01:33.760
AI governance platform that gives continuous
00:01:33.760 --> 00:01:37.900
behavioral understanding of agentic AI. The top
00:01:37.900 --> 00:01:40.920
10 finalists each walked away with a 5 million
00:01:40.920 --> 00:01:44.219
investment commitment. The money is clearly following
00:01:44.219 --> 00:01:47.840
agent observability and agent governance. And
00:01:47.840 --> 00:01:51.420
on the offensive side, Mendean's Amtron's 2026
00:01:51.420 --> 00:01:54.060
report confirmed what Talos and others have been
00:01:54.060 --> 00:01:57.260
whispering about. Adversaries have moved from
00:01:57.260 --> 00:02:00.719
experimental AI use. They're now deploying autonomous
00:02:00.719 --> 00:02:03.739
agents that rewrite their own code in real -time.
00:02:04.340 --> 00:02:07.200
High -prose disclosure of operation Ola Lampo
00:02:07.200 --> 00:02:11.930
and Rust -based Charbactor was. To quote one
00:02:11.930 --> 00:02:15.169
analyst, the sharpest signal that AAS system
00:02:15.169 --> 00:02:18.310
malware development by state sponsor actors is
00:02:18.310 --> 00:02:22.889
no longer theoretical. Let's talk concrete announcements.
00:02:23.050 --> 00:02:25.370
I'll group them by what problem they are actually
00:02:25.370 --> 00:02:29.189
trying to solve. This was the largest category
00:02:29.189 --> 00:02:32.810
by volume, the agent identity and runtime control.
00:02:33.789 --> 00:02:36.110
Cisco announced defense glue, an open source
00:02:36.110 --> 00:02:39.560
secure agent framework. and extended zero trust
00:02:39.560 --> 00:02:45.180
access to AI agents through dual IAM so you can
00:02:45.180 --> 00:02:47.879
assign time bound. It has specific permissions
00:02:47.879 --> 00:02:51.240
to an agent the same way you would a contractor.
00:02:52.419 --> 00:02:55.400
Microsoft Shift and new identity features in
00:02:55.400 --> 00:02:58.580
the entry ID to track and control agent behavior.
00:02:59.300 --> 00:03:02.759
Plus guardrails previews in Microsoft Foundry.
00:03:03.280 --> 00:03:06.520
Teleport launch beams. Each agent runs in an
00:03:06.520 --> 00:03:10.280
isolated firecracker micro -VM with its own identity
00:03:10.280 --> 00:03:13.419
connected to infrastructure without secrets,
00:03:13.800 --> 00:03:18.699
full audit trail per agent. One password launch
00:03:18.699 --> 00:03:22.020
unified access for credential and machine identity
00:03:22.020 --> 00:03:26.139
control, with Anthropic, Cursor, GitHub, Perplexity,
00:03:26.340 --> 00:03:29.840
and Vercell as launch partners. Astric Security
00:03:29.840 --> 00:03:32.979
expanded to cover shadow agent deployments on
00:03:32.979 --> 00:03:36.900
managed devices. agentic sock and the death of
00:03:36.900 --> 00:03:41.560
the seam. Databricks CEO Ali Goldzi told CNBC
00:03:41.560 --> 00:03:45.099
and on the show floor quote AI will kill the
00:03:45.099 --> 00:03:49.280
seam in 2026. Bo claimed the product behind it
00:03:49.280 --> 00:03:53.340
Lakewatch an agentic seam powered by Claude targeted
00:03:53.340 --> 00:03:57.860
squarely at sock infrastructure. Databricks also
00:03:57.860 --> 00:04:01.240
extend its AI security framework with 35 new
00:04:01.240 --> 00:04:05.129
technical risks and mitigation controls specifically
00:04:05.129 --> 00:04:09.590
for agent memory and planning, which tells you
00:04:09.590 --> 00:04:12.569
where they think the real attack surface lives.
00:04:13.590 --> 00:04:16.089
Google Cloud rolled out a new agentic automation
00:04:16.089 --> 00:04:20.050
in Google security operations, including a tree
00:04:20.050 --> 00:04:23.850
edge and investigation agent that autonomously
00:04:23.850 --> 00:04:27.870
gathers evidence and returns verdict with explanations.
00:04:29.110 --> 00:04:33.230
Accenture and Entropic launched Cyber AI, Cloud
00:04:33.230 --> 00:04:36.629
as the reasoning engine for soft workflow automation
00:04:36.629 --> 00:04:40.350
with a library of autonomous agents across the
00:04:40.350 --> 00:04:45.490
security lifecycle. CrowdStrike unveiled Charlotte
00:04:45.490 --> 00:04:49.269
AI
Load full transcript
AgentWorks, a no -code platform and ecosystem
00:04:49.269 --> 00:04:53.250
for building custom security agents on Falcon.
00:04:54.350 --> 00:04:57.610
Launch partners include AWS, Anthropic, NVIDIA,
00:04:57.769 --> 00:05:01.399
OpenAI, and Salesforce. Sentinel -1 announced
00:05:01.399 --> 00:05:05.019
prompt AI agent security, a real -time governance
00:05:05.019 --> 00:05:08.879
control plane for unauthorized agentic actions.
00:05:09.680 --> 00:05:12.240
AI generated code security. This one matters
00:05:12.240 --> 00:05:15.480
for anyone shipping code with copilot or cursor.
00:05:16.420 --> 00:05:20.560
Secure Code Warriors Trust Agent AI gives commit
00:05:20.560 --> 00:05:24.139
level visibility into which LLM influence which
00:05:24.139 --> 00:05:27.180
commit. Correlates that with the vulnerability
00:05:27.180 --> 00:05:31.139
exposure and tracks. Active MCP servers to stop
00:05:31.139 --> 00:05:35.839
agents from breaching unvetted tools. Block .doc
00:05:35.839 --> 00:05:39.120
.signal is an agentic abstract product, purpose
00:05:39.120 --> 00:05:42.040
-built for autonomous development workflows.
00:05:43.040 --> 00:05:46.600
Post -quantum quietly, overshadowed by AI but
00:05:46.600 --> 00:05:51.279
moving fast because NSA CNSA 2 .0 deadlines are
00:05:51.279 --> 00:05:55.100
real. Zero -tier, launch its quantum platform
00:05:55.100 --> 00:06:00.399
with hybrid PQC builds. into the transport protocol,
00:06:00.920 --> 00:06:03.600
architecturally one of the most interesting moves
00:06:03.600 --> 00:06:07.639
at the show. Palo Alto Networks added next -generation
00:06:07.639 --> 00:06:11.620
trust security to automate certificate lifecycle
00:06:11.620 --> 00:06:18.600
and PQC readiness. PQCEE shipped a crypto -agile
00:06:18.600 --> 00:06:22.240
CNG provider for Windows, letting you swap in
00:06:22.240 --> 00:06:25.399
post -quantum algorithms at the OS crypto layer.
00:06:26.089 --> 00:06:31.389
Sandbox AQ expanded AQ active guard for enterprise
00:06:31.389 --> 00:06:37.569
PQC migration in regulated industries Palo Alto's
00:06:37.569 --> 00:06:40.509
full stack worth calling out separately Prisma
00:06:40.509 --> 00:06:45.230
Ares 3 .0 to secure the full agentic AI lifecycle
00:06:45.230 --> 00:06:50.829
agentic SASE enhancements to Prisma SASE and
00:06:50.829 --> 00:06:54.290
Prisma browser for business they are going wide
00:06:55.880 --> 00:06:58.920
The cryptographers panel and what's underneath.
00:06:59.399 --> 00:07:02.459
The 25th anniversary cryptographers panel had
00:07:02.459 --> 00:07:06.060
a quote from the moderator Paul Kocher. That's
00:07:06.060 --> 00:07:08.720
worth sitting with. The cryptographic algorithms
00:07:08.720 --> 00:07:12.360
have really held up over the last 25 years and
00:07:12.360 --> 00:07:14.579
you can say that about a lot of areas within
00:07:14.579 --> 00:07:17.759
cybersecurity. Think about that. The math is
00:07:17.759 --> 00:07:20.639
holding. The implementations are holding. What's
00:07:20.639 --> 00:07:23.980
not holding is everything we bottle on top. Identity,
00:07:24.079 --> 00:07:27.319
access control. that pretty much model human
00:07:27.319 --> 00:07:31.060
speed sock operations, the industry is not having
00:07:31.060 --> 00:07:34.439
a cryptographic crisis, it's having an architecture
00:07:34.439 --> 00:07:38.720
for autonomous actors crisis. That's why the
00:07:38.720 --> 00:07:41.300
Jericho's forum and BeyondCorp concepts keep
00:07:41.300 --> 00:07:44.660
coming back at the RSA conference. Zero Trust
00:07:44.660 --> 00:07:49.339
is 15 plus years old as a concept and most enterprises
00:07:49.339 --> 00:07:52.560
still have not implemented it for human users,
00:07:52.939 --> 00:07:56.620
let alone for agents. Cisco even ran a panel
00:07:56.620 --> 00:07:59.879
called Despite years of enthusiasm since the
00:07:59.879 --> 00:08:02.939
Jericho forum and Google's Beyond Court, many
00:08:02.939 --> 00:08:06.240
still struggle to implement Zero Trust. That
00:08:06.240 --> 00:08:12.000
is an honest subtitle. What is actually coming?
00:08:12.720 --> 00:08:14.920
Here's what I would tell any CISO or senior engineer
00:08:14.920 --> 00:08:17.480
to pay attention to over the next 12 months,
00:08:18.180 --> 00:08:21.370
based on what I saw and read from the show. 1.
00:08:21.529 --> 00:08:24.550
Non -human identity becomes the primary identity
00:08:24.550 --> 00:08:27.970
problem. If you're still treating service accounts,
00:08:28.449 --> 00:08:31.170
API keys, and agents as an afterthought to human
00:08:31.170 --> 00:08:34.470
IAM, you have maybe 6 months before this becomes
00:08:34.470 --> 00:08:37.929
an audit binding. Look at one password unified
00:08:37.929 --> 00:08:41.289
access, asterisks, and how Entra and Okta are
00:08:41.289 --> 00:08:45.450
extending agent governance. 2. MCP is now part
00:08:45.450 --> 00:08:48.620
of the attack surface. The model context protocol
00:08:48.620 --> 00:08:52.740
is how agents get tools. Stitch, now Twilio -owned
00:08:52.740 --> 00:08:56.019
and Greylock, shipped MCP -specific monitoring
00:08:56.019 --> 00:08:59.860
at RSA conference. If your org is building agents,
00:09:00.080 --> 00:09:02.860
you need an MCP gateway strategy, full stop.
00:09:03.460 --> 00:09:06.860
Treat MCPs like you treat API gateways, inventor
00:09:06.860 --> 00:09:10.100
it, authenticate it, log it and rate limit it.
00:09:10.519 --> 00:09:14.419
The SOC gets automated or gets outrun. 22 seconds
00:09:14.419 --> 00:09:17.559
handoff times are not something that tier 1 analysts
00:09:17.559 --> 00:09:21.340
can chase. Agentic triage is no longer optional.
00:09:21.779 --> 00:09:24.419
It's the only way to get mean time to respond
00:09:24.419 --> 00:09:28.039
below the attacker's mean time to pivot. Watch
00:09:28.039 --> 00:09:31.080
Google SecOps, Lakewatch, and CrowdStrike Charlotte.
00:09:32.000 --> 00:09:35.399
4. Prompt injection is the new SQL injection.
00:09:36.179 --> 00:09:38.860
Veeam's Rihal Jalil put it bluntly at the RSA
00:09:38.860 --> 00:09:42.370
conference. Every day there's news of agents
00:09:42.370 --> 00:09:46.090
doing something funky with enterprise data. OWOS
00:09:46.090 --> 00:09:49.909
has the LLM top 10 but the real engineering question
00:09:49.909 --> 00:09:53.549
is where in your architecture do you do content
00:09:53.549 --> 00:09:56.570
level inspection on agent inputs and outputs.
00:09:57.870 --> 00:10:02.870
Bonfee ACS 2 .0, Rubrik Sage, and Sanity all
00:10:02.870 --> 00:10:07.389
took a crack at this. 5. Post quantum is closer
00:10:07.389 --> 00:10:11.820
than you think. CNSA 2 .0 timelines mean that
00:10:11.820 --> 00:10:15.299
anything you procure now with a 5 plus year hardware
00:10:15.299 --> 00:10:19.659
life cycle needs a PQC story. This is not a 2030
00:10:19.659 --> 00:10:24.320
problem anymore. Start inventorying. 6. Active
00:10:24.320 --> 00:10:28.480
defenses and deception are coming back. A Calvius
00:10:28.480 --> 00:10:32.059
360 deception is explicitly designed to break
00:10:32.059 --> 00:10:36.019
AI driven attack automation. When attackers automate
00:10:36.330 --> 00:10:39.610
automate reconnaissance with agents, deception
00:10:39.610 --> 00:10:42.730
scales disproportionately well against them.
00:10:43.190 --> 00:10:47.450
Accept more of this. Now to close, the honest
00:10:47.450 --> 00:10:51.730
takeaway from RSA conference 2026. The industry
00:10:51.730 --> 00:10:55.070
has finally stopped pretending agentic AI is
00:10:55.070 --> 00:10:58.730
a future problem. Every major vendor ships something
00:10:58.730 --> 00:11:02.299
real. Some of it will stop ship as vaporware,
00:11:02.779 --> 00:11:05.220
some of it will be absorbed by acquisition before
00:11:05.220 --> 00:11:09.299
it matures, but the category is here. The uncomfortable
00:11:09.299 --> 00:11:13.279
line from Illumnius Carl Van Den Berg is the
00:11:13.279 --> 00:11:16.259
one I keep coming back to. We're spending more
00:11:16.259 --> 00:11:19.899
on cybersecurity than ever before, but the breaches
00:11:19.899 --> 00:11:23.240
keep happening. Something is fundamentally broken
00:11:23.240 --> 00:11:27.190
about how we have approached this problem. If
00:11:27.190 --> 00:11:30.409
22 seconds is the new hand of time and agents
00:11:30.409 --> 00:11:33.129
are now both the most productive workers and
00:11:33.129 --> 00:11:35.570
the most dangerous insider in your environment,
00:11:36.129 --> 00:11:39.169
then the next 12 months are about to walk. About
00:11:39.169 --> 00:11:42.570
one question. Can you operate your security program
00:11:42.570 --> 00:11:45.730
at machine speed without losing human accountability?
00:11:46.350 --> 00:11:49.529
If you can, the agentic era is an advantage.
00:11:49.830 --> 00:11:53.669
If you can't, it's an extinction event. That's
00:11:53.669 --> 00:11:56.529
the episode. This has been Tech Update. Go break
00:11:56.529 --> 00:11:58.669
something in a lab today so you don't have to
00:11:58.669 --> 00:12:00.450
explain it in production next week.