Cyber + AI + Community, Issue 2

Do you agree that deep cyber experience, practical AI engineering, and an active open community are crucial to staying ahead of modern attackers? This is edition 2 of CAC, our Cyber, AI and Community newsletter. Please send us any feedback, and share this with colleagues if you find it valuable.

An NSA founded OT cyber center endorses DeepTempo

For the last several months, DeepTempo has been collaborating with the Technology Advancement Center. Established by the National Security Agency, and collaborating today with the Defense Information Systems Agency and other national and state national security organizations, the TAC is a leader in evaluating and disseminating approaches to operational technology cyber security. We have published a case study with the TAC demonstrating the use of Vigil and our purpose built LogLM to identify and respond to increasingly advanced attacks on critical infrastructure.

If you defend critical infrastructure, the case study shows what behavioral detection caught in a live OT environment, and what it cost to run.

Read here

Branch61G is our weekly read on what actually hit, written from the grass rather than the broadcast booth.

The Sleeper Picks: When BGP Hijacks Poison the Patch Pipeline

Josiah Lashley · September 2

Between August 28 and August 30, an unrelated network (AS62390) announced a more specific route (162.55.80.0/24) for Hetzner IP space hosting Softaculous and Virtualizor update infrastructure. BGP preferred the route, diverting traffic intended for global hosting patch and billing servers to attacker-controlled nodes. The attackers pulled valid TLS certificates for affected domains, so downstream hypervisors checking for updates accepted the connection without a single browser or transport warning, pulling backdoored packages that installed a persistent /etc/systemd/system/java-jre-update.service unit, rogue SSH keys, and cron jobs.

Josiah's takeaway: the enterprise perimeter is no longer just appliances on your rack; it is the upstream trust chains feeding them. When an adversary controls the route and the certificate, standard endpoint defenses assume compliance. Watching anomalous outbound beaconing and lateral movement at the behavioral level is often the only tripwire left.

Read here

Two previous B61G dispatches dropped since our first newsletter:

Week 5 (Aug 25): Three dogs, a live range, and the channels that already belong. Living-off-the-land zero-days added to CISA's KEV, including the Zimbra SNMP unauth RCE (CVE-2026-73570) and the Oracle WebLogic proxy access bypass (CVE-2026-21962), put into TempoRange to test LogLM's behavioral detection against live lateral movement. Read it

Week 4 (Aug 18): The crew you trust, and the control planes attackers already live in. A China nexus operator turning a VMware vCenter syslog flaw into root, then hiding behind cron jobs that impersonate VMware itself, plus SAP Commerce Cloud exploitation, a GeoServer SQL injection, and Clop's PLM data theft claims. Read it

Tokenomics Begins Upstream

Evan Powell · August 31

Shaped by extensive feedback from tier-one operators and MSSPs, Evan tackles the reality that as attackers weaponize AI, they can "economically DDoS" defenders by flooding environments with polymorphic noise to exhaust token budgets on frontier models. No contract term fixes this. The architecture does. It begins upstream, by running encoder-only models like LogLM directly alongside the data lake or pipeline where telemetry lives, rather than paying exorbitant ingestion taxes to ship raw logs to a SaaS SIEM.

From there, a token-efficient harness like Vigil evaluates the investigative state, letting fast local models resolve 95%+ of routine triage and escalating only when the model hesitates or evidence crosses corroboration thresholds. Static procurement estimates are replaced with continuous, empirical tuning tied to actual threat volume.

Read here

Agent Skills and Ice Houses: Notes from San Francisco

Mayank Kumar · August 28

On August 26, the AI Cyber Alliance gathered at Workato's headquarters in San Francisco for an in-person, practitioner-only session with zero vendor pitches. The discussion centered on two converging primitives: building modular, schema-constrained Agent Skills via the Model Context Protocol (MCP) instead of stuffing bloated instructions into system prompts, and standing up Ice Houses (Apache Iceberg plus Trino and data lakehouses) for security telemetry. Retaining months of raw, structured telemetry at commodity storage costs gives upstream foundation models like LogLM the high-fidelity historical data needed for autonomous investigations without vendor lock-in.

Read here

Vigil is the leading open source AI SOC, Apache 2.0, built in the open by the community.

Lessons Learned Building a Harness for Cyber Security

Mayank Kumar · August 27

Drawing on experience building foundation models at the Allen Institute and automation engines like StackStorm, this deep dive explains why Vigil abandoned fragile SOAR playbooks and endless prompt-tweaking. The core thesis: a prompt is not an instruction, it is a coordinate. Inference weights are frozen; a prompt merely selects existing capability. Vigil frames the harness as a deterministic selection function where a Python controller runs a fixed state machine with a closed schema (INVESTIGATE, EXPAND, PIVOT, DEEPEN, ABANDON, VALIDATE, HANDOFF_IR, CHECKPOINT, CONCLUDE).

The model recommends; the controller decides. Gating relies on verifiable evidence_strength rather than self-reported confidence, and prompt injection in logs, for example fake pentest claims arguing for ABANDON, is neutralized in code: code may promote salience, only humans may demote, and self-exonerating telemetry can never trigger an exit.

Read here

VigilSOC 0.5.0 Release: Upgrades, Governance, and Office Hours

John Van Lowe · August 20

VigilSOC 0.5.0 is live on GitHub, featuring optimized event processing pipelines that slash latency under high load, an expanded telemetry rule engine, streamlined CLI and APIs, and major stability fixes. The release also formalizes our repository discipline: a pull-based queue with a strict WIP cap of 10 open issues per milestone, a 15-label taxonomy (P0 to P3 urgency), explicit Definition of Done standards, and formalized security reporting mechanisms.

A massive thank you to our 0.5.0 contributors: craig-dt, cyforkk, G-r-ay, jmurph1, josiahlashley21, joshuacox, mattmorr1, mgalore, nestor-deeptempo, pollychen-lab, samarmstrong, thetosy, and tomatotomata. Community Office Hours are now running regularly on Discord.

Read here
GitHub: github.com/Vigil-SOC/vigil
Discord: discord.gg/SBtSHzMYFZ
Benchmarks: socbench.org

Everyone welcome: transparent nomination for AI Cyber Alliance speakers and hosts

You might have heard about a community for technical conversations and demonstrations of AI and Cybersecurity called the AI Cyber Alliance. Recently the AI Cyber Alliance established ongoing chapters in San Francisco, Austin, Boston, DC (Northern Virginia), and the SF Bay Area. There is now enthusiasm for these and other communities to organize themselves, much like DevOps days became a local phenomenon across the United States and around the world. With that in mind, the AI Cyber Alliance website and underlying GitHub is evolving towards self organization. If you or someone you know would be a great speaker at an AI Cyber Alliance meeting, please take a look at the new site and get involved. You can also nominate yourself to host an event.

Read here

AI powered attackers have already transformed themselves. How are we going to do the same? Hype based superficial marketing is unlikely to help. Technical, real world, no BS conversations are required for useful patterns to be rapidly discovered and shared. If you agree, the AI Cyber Alliance is for you.

Cyber defense must move with more precision and intelligence than attackers. In our next edition, look for:

• Additional VigilSOC architectural updates covering multi-agent parallel operations and environmental memory layers.
• Full recaps from our first Austin AI Cyber Alliance gathering.
• Dates and RSVP links for upcoming fall Alliance events across Boston, the Bay Area, and San Francisco.

Tell us what you would like to see covered, and subscribe if you would like to receive updates as we publish additional CAC releases.