Express YourSelf !

Reference https://www.meetup.com/collabnix/events/316301263/

1. Contain, Forecast, Adjudicate: Three Agent Controls No Protocol Gives You

https://www.nasiko.com/

2. Securing Your Agentic Stack (Workshop) 

https://agentic.dockerworkshop.com/

https://agentic.dockerworkshop.com/#/labs/securing-the-agentic-stack-slides

https://agentic.dockerworkshop.com/#/labs/securing-the-agentic-stack

1. What it contains? What is software Artifact

SBOM

docker scout sbom --format spdx --output baseline.spdx.json catalog-service:baseline

2. From where it comes from?

Provenience

3. CI pipeline. Who approve image with vulnerabilities? Can I. verify attestation source

https://docs.docker.com/scout/deep-dive/advisory-db-sources/

match with SBOM

Now AIBOM term is emerging. https://www.ajeetraina.com/ai-bom-explained-why-your-sbom-stops-where-your-ai-system-starts

VEX: Vulnerability Exploitability eXchange 

In Docker Desktop, we can filter vulnerabilities based on fixable or not

SLSA

Level 1 to 3

FIPS 140 For US

4. Can it be restricted? Sandbox  

We need base image with near zero vulnerabilities.

All images shall be signed

Allow coding agent only what it needs.

"/rc" in Claude. Remote control. You will keep getting notifications in your mobile. 

AI Agent

local sandbox and cloud sandbox. SBX is running on microVM. Agent can change kernel also. So SBX runs on microVM instead of container.

Inside mircoVM also we run Docker engine and docker daemon 

Docker Hub have MCP policy, AI policy

Now we have SBX compose file with policy

MCP Toolkit

Hardened MCP servers

https://agentic.dockerworkshop.com/#/labs/securing-the-agentic-stack-slides/workshop-75

Docker hub

DHI Docker Hardening Images

Images and AI models also on 

3. Docker sbx kits: you explorations to contain AI Agents begins here

AI Agent

https://docs.docker.com/ai/sandboxes/customize/

https://docs.docker.com/ai/sandboxes/customize/kits/

https://hub.docker.com/search?type=sbx_kit

Tools

1. MIXIN kit

It has enhanced capabilities

build agent from scratch

2. Sandbox Kit

-------

Kit has spec.yaml file

Files are payload. it can have docker compose file. certificate file etc. Some will go to sandbox and other files remain on laptop

start from Mixin kit, as Sandbox kit has many definition

Now let's have customize AI agent. 

https://floci.io/ is like localstack. Cloud emulators 

floci CLI is inside sandbox

1. create shell sendbox

2. run docker compose

same can be done with spec.yaml file

We have DHI for langchain also. It can be inside sandbox

4. Beyond the Agent: Building AI Systems You Can Trust

If it hallucinate then workflow has problem or model has problem? 

5. The New Primitives of AI: YAML, OCI, and Agent Infrastructure

you write agent in your Jupiter Notebook

"It works on my notebook"

"docker agent"

oci artifacts

I did not attend this event in-person. I gone through YouTube Playlist https://www.youtube.com/playlist?list=PLcWDDGrTp5AU It has 41 videos

In part 1, let me cover few of them

-------------------

1. Made in India. the founders behind API tools

• specmatic
• Beeceptor
• keploy.io
• postman
• bruno
• karate labs

------------------------

2. Restoring Trust in AI-Native development

Earlier we used to have Pre-commit hook before agent started coding. 

Now, SDD = spec driven development

Vibe coding was based on prompt

now spec is new source code

Harness engineering

1. Guides gives feedback to agent

2. Sensors for self-correcting loop

3. Executable intent

4. Executable architecture

5. Continuous Governance

------------------------

3. Death of API Management

The speaker is founder of "bunny and cloud" a collaborative development tool for humans and AI

https://bunnyandcloud.com/

He explain reasons. 

1. A New practices is emerging: Context engineering

A right context at right time for agent, so it can reason and act reliably. 

2. Agents are taking all management attention

AI GWs and context contracts extend API Management into agent governance. 

3. API Management does not get intent

But agent needs intent

We need to hard code and orchestrate the agents for different workflows. Agent does not think about workflow. 

4. APIs are relegated to the execution layer

Reasoning layer by LLM (Probabilistic) 

context layer by MCP and RAG

Execution layer (deterministic) 

5. The GW shift

Kong is decoupling API GW and AI GW

Portkey AI GW pioneer is acquired by Palo Alto Networks

AI GW enforce policy to every call to LLM, MCP, RAG

6. The tokenomics is replacing APInomics

7. Context Management includes managing APIs

Microcontext. Not DB dump. Agent shall receive smallest truthful slice of bounded context. 

8. From API endpoint Management to capabilities management

Agent Registries

Instead of DX, now we need AX (Agent eXperience) 

Context Management is new API Management

pillar 1: Identity and intent context

pillar 2: Business and domain context

pillar 3: Knowledge and evidence context (RAG) 

pillar 4: Execution and feedback context (MCP)

AI GW examples

1 Portkey AI GW

2 Kong

3 Truefoundry (someone added from the audience) 

The attack surface is different for AI GW. 

36. Skills and MCP

https://www.skills.sh/

API related classes / certifications

https://apimasters.io/

 Kyverno Policy

- 2 types: Policy and ClusterPolicy

- Multiple Rules

- Match / Exclude

-- Match resources

kind is mandatory. names, namespaces, operations, selector are optional. 

Wildcards * supported in kinds, names, namespaces

All are AND condition. 

Any means OR condition

We can mention based on who created

exclude:

any:
- clusterRoles:
- cluster-admin
- subjects:
- kind: User
name: John

match AND exclude

exclude must be a subset of match

- Action

1. Validate ( allow / deny )

2. Mutate

3. Generate new K8s object

4. Verify image (Cosign/Sigstore)

- Enforce / Audit

* It uses JAMESPath

- Filter JSON

# Returns a list of container objects that match the condition
{{ request.object.spec.containers[?starts_with(image, 'nginx')] }}

validate:

message: "Nginx images are not allowed!"
deny:
conditions:
all:
# Filter the list. Use length() to count.
# If count > 0, it means a violation exists -> Block.
- key: "{{ request.object.spec.containers[?starts_with(image, 'nginx')] | length(@) }}"
operator: GreaterThan
value: 0

- pipe

Used to extract a specific field from a complex object into a flat list.

# Input: List of container objects
# Output: ["nginx:latest", "redis:alpine", "busybox"]
key: "{{ request.object.spec.containers[].image }}"

kyverno jp query -i object.json 'spec.containers[].name'

Combo (Filter + Pipe): “Get the containerPort of the container named ‘app’”

 request.object.spec.containers[?name == 'app'].ports[].containerPort

You want to find the names of all volumes that are of type emptyDir.

spec.volumes[?emptyDir != null].name

kyverno jp query -i object.json "spec.volumes[?emptyDir != null]"

- null handling

Risk: {{ request.object.metadata.labels.team }} (If null -> Error).

Safe: {{ request.object.metadata.labels.team || '' }} (If null -> treat as empty string).

- The length function

Example: “A Pod must not have more than 3 containers.”

validate:
deny:
conditions:
all:
- key: "{{ request.object.spec.containers | length(@) }}"
operator: GreaterThan
value: 3

- if / else 

- sum

- contain means exist : contains(request.object.metadata.labels, 'production')

advance mutate foreach

mutate:

foreach:
- list: "request.object.spec.containers"
patchStrategicMerge:
spec:
containers:
- name: "{{ element.name }}"
securityContext:
readOnlyRootFilesystem: true

* Aut-Gen: If rule for pod then automatically generate rules for Deployment, StatefulSet, DaemonSet, etc.

https://release-1-8-0.kyverno.io/docs/writing-policies/autogen/

Generate rule has synchronize flag

synchronize: true: Kyverno complete managed lifecycle

synchronize: false: Kyverno created for the first time then user can edit it manually without getting revert back like synchronize: true

* request.object is the incoming resource configuration (the new state) that is being submitted to Kubernetes API server

Documentation: https://kyverno.io/docs/policy-types/cluster-policy/variables/ 

  # -------------------------------------------------------------

  # ACTION: If Kyverno is dead, just let the request go through.

  # -------------------------------------------------------------

  failurePolicy: Ignore # or Fail

PolicyReport

PolicyReport stores the results of those rules.

The PolicyReport is essentially a “Health Check Report Card” for your Kubernetes resources.

Its main goal is Observability & Auditing.

It provides summary and then detail about all failures. For Example

apiVersion: wgpolicyk8s.io/v1alpha2
kind: PolicyReport
metadata:
name: polr-ns-default
namespace: default # It lives next to the Pod, not at the cluster level
labels:
app.kubernetes.io/managed-by: kyverno
summary:
pass: 0
fail: 1
warn: 0
error: 0
skip: 0
results:
- policy: require-labels # The name of the ClusterPolicy responsible
rule: check-for-team-label # The specific rule name
category: Best Practices
severity: medium
result: fail # The outcome (fail, pass, warn, error, skip)
message: "Validation error: label 'team' is required"
source: kyverno
resources: # The specific object that failed
- apiVersion: v1
kind: Pod
name: nginx
namespace: default
uid: a1b2c3d4-e5f6...

* Background scan never delete resource even with enforce mode

* For background scan following variable are not relevant

request.userInfo.*

request.operation

request.dryRun

serviceAccountName in admission context

Document: https://kyverno.io/docs/policy-reports/background/

* Cleanup policy delete pods

https://kyverno.io/docs/policy-types/cleanup-policy/

Kyverno CLI

* for given resource, policy is pass or fail

kyverno apply policy.yaml --resource pod.yaml

* test JAMESPath expression against JSON

kyverno jp query -i object.json 'metadata.labels'

* Check policy YAML is written correctly or not

kyverno validate policy.yaml

External Data Source

Purpose: It allows you to load data from outside into a variable before the rule logic (validate/mutate) runs.

In order to consume data from a ConfigMap in a rule, a context is required... The context data can then be referenced in the policy rule using JMESPath notation.

Kyverno supports 3 main data sources in context:

1. Kubernetes Resources (via API Call): Look up existing data in the cluster (e.g., ConfigMaps, Secrets, Services).

2. External APIs: Make an HTTP call to a service outside the cluster.

3. Image Registry: Fetch metadata about a container image (e.g., image size, architecture).

verifyImages:
- imageReferences:
- "ghcr.io/myorg/*"
attestors:
- entries:
- keys:
publicKeys: |-
-----BEGIN PUBLIC KEY-----
...
-----END PUBLIC KEY-----

https://main.kyverno.io/docs/policy-types/cluster-policy/external-data-sources/

Kyverno Mutate — JSON Patch

the standard patchStrategicMerge merges YAMLs together.

patchesJson6902 is a mutation method specific operations (RFC 6902 standard) to tell Kyverno exactly how to change the data.

Usage: 

* Removing a field (impossible with standard merge).

* Adding an item to a specific position in a list (arrays).

* Replacing a value entirely without merging.

It follows the JSON Patch format:

* op: The action (add, remove, replace).

* path: The location of the field (e.g., /metadata/labels/mytag).

* value: The data to put there.

Example: “Add a sidecar container”

https://main.kyverno.io/docs/policy-types/cluster-policy/mutate/#rfc-6902-jsonpatch

TTL Default of Certification in Kyverno:

https://pkg.go.dev/github.com/kyverno/kyverno/pkg/tls

const (

    CAValidityDuration = 365 * 24 * time.Hour      // 365 days

    TLSValidityDuration = 150 * 24 * time.Hour     // 150 days

    CertRenewalInterval = 12 * time.Hour           // 12 hours

)

https://www.udemy.com/course/complete-certified-kyverno-associate-kca-exam-prep/

https://medium.com/@kienlt.qn/prepare-for-the-kyverno-certified-associate-kca-exam-c144906f9bc2

Concurrent Policies Generation Number Default!

PolicyException CRD

CEL

RNN. Seq to seq NLP tasks. 

1. Many to one: Sentimental Analysis

2. One to Many: Image caption

3. Many to Many: 

- Synch many to many: # input = # output. E.g. Part of speech tagging, Named Entity Recognition

- Asynch many to many: translation, text summarization, question and answer, chatboat, speech to text, 

Seq2seq model is used for Many to Many

Stage 1: 2014 Encoder decoder network

Encoder and decoder are LSTM. RNN and GRU are other options. 

It is good for small sentences. Not for 30+ words

BLEU score

Stage 2: 2015 Attention Mechanism

Encoder is same

Attention Mechanism: Attention layer at decoder finds out which hidden state is useful at each stage of decoder and generate context vector for that stage. So, Multiple context vectors based on  encoder's (hidden state of LSTM = ctht vector) are available to decoder. 

Training time is more. 

2015 to 2017: May types of Attention Mechanisms were introduced. 

Stage 3: 2017 Transformer

No LSTM

No RNN Cell

Self-attention was introduced

Both encoder and decoder uses attention

Transformer can process all words in parallel 

1. Attention layer = Multi Head Attention

2. Normalization Layer

3. Dense Layer

4. Input embeddings

It needs hardware, time, and data

Stage 4: 2018 Jan Transfer Learning

Challenges

1 Single model cannot perform all tasks like sentimental, translation, summarization

2 lots of labeled data

Universal Language Model Fine-tuning ULMFiT proposed to use Language modelling  as Pre-training. Language modelling is NLP task to predict next word. Advantages

1. Rich feature training

2. unsupervised task

model: AWD LSTM model

data set: wikipedia

finetuning changed output as classifier with many data set 

Scratch 10000 data. Now fine tune 100 data still better result

- No transformer

Now in 2018, we have two technolgoies

1. architecture: transformer

2. training. Pretrain and transfer learning

Stage 5: 2018 Oct LLM

Transfer learning on transformer

1. Google : BERT (encoder only model) 

2. OpenAI: GPT (decoder only model)

LM to LLM

1. data

2 hardware GPU clusters

3 time : days to weeks

4. cost =  h/w + electricity + people + infra

5. energy consumption 

---------------

GPT3 - > chatGPT

1. RLHF : Reinforcement Learning from Human Feedback

2. incorporate safety and ethical guidelines 

3. improvement in contextual point

4. dialogue specific 

5. continuous improvement based on user feedback

Reference https://www.youtube.com/watch?v=8fX3rOjTloc&list=PPSV