Agentic Testari Sentinel Hub · FriendlyAI agentictestari.com
Confidential
L3 MAX Public-max · non-destructive

Public surface assessment — CGI

Enterprise marketing edge. Strong email; CSP quality is the main HIGH class (unsafe-inline/eval, polyfill.io, huge third-party allowlist) plus large internal-style DNS inventory. True L3: HIGH 3 · MEDIUM 3.

Target cgi.com Date 2026-07-29 RoE public-max · non-destructive Engagement eng_cgi_l3_1358
3
HIGH
3
MEDIUM
0
INFO
OK

Executive summary

Why act now

These findings are not theoretical: they are public surfaces a human or a hacker agent can consume today. Each card has proof links, impact, owner, fix, and validation — so the team takes risk seriously and acts.

Each finding includes: problem, where (clickable link), attacker/agent impact, owner, PoC (read-only), how to fix, and how to validate.

Enterprise marketing edge. Strong email; CSP quality is the main HIGH class (unsafe-inline/eval, polyfill.io, huge third-party allowlist) plus large internal-style DNS inventory. True L3: HIGH 3 · MEDIUM 3.

Stack in this run

ComponentUsedNote
ParviClaw Core + AWPYES (lab)full pack on engagement
ParviSight browserpartialHTTP walk / agentic
PayBotFin Witness liveas engagedhonest scope in annex

Findings

Each finding includes: problem, where (clickable link), attacker/agent impact, owner, PoC (read-only), how to fix, and how to validate.

CG-H1

CSP unsafe-inline + unsafe-eval

HIGH
The problem

https://cgi.com/ response headers — Production script-src includes 'unsafe-inline' and 'unsafe-eval' plus a very large host allowlist. last_reprobe=2026-07-31.

Weak browser security headers increase impact of XSS/clickjacking/MITM.

What an attacker or hacker agent can do
  • Human: Public headers surface on cgi.com expands attacker recon.
  • Agent (LLM+tools): Keep surface in continuous recon.
  • If invasive/destructive: Escalate if second bug/credential. L3 public-max is read-only.
Who owns it

CGI Inc. — security / product / infrastructure (confirm internal owner)

PoC (proof — does not fix)

Read-only. Does not change the system.

curl -sSI 'https://cgi.com/' | egrep -i 'strict-transport|content-security|x-frame|x-content'
How to fix

Add/tighten HSTS, CSP, XFO, nosniff, Referrer-Policy, Permissions-Policy on https://cgi.com/.

How to validate the fix

Add/tighten HSTS, CSP, XFO, nosniff, Referrer-Policy, Permissions-Policy on https://cgi.com/.

CG-H2

CSP trusts polyfill.io + broad CDNs

HIGH
The problem

script-src includes https://polyfill.io (and other broad CDNs like unpkg, rawgit, jsdelivr). last_reprobe=2026-07-31.

Weak browser security headers increase impact of XSS/clickjacking/MITM.

What an attacker or hacker agent can do
  • Human: Public headers surface on cgi.com expands attacker recon.
  • Agent (LLM+tools): Keep surface in continuous recon.
  • If invasive/destructive: Escalate if second bug/credential. L3 public-max is read-only.
Who owns it

CGI Inc. — security / product / infrastructure (confirm internal owner)

PoC (proof — does not fix)

Read-only. Does not change the system.

curl -sSI 'https://polyfill.io' | egrep -i 'strict-transport|content-security|x-frame|x-content'
How to fix

Add/tighten HSTS, CSP, XFO, nosniff, Referrer-Policy, Permissions-Policy on https://cgi.com/.

How to validate the fix

Add/tighten HSTS, CSP, XFO, nosniff, Referrer-Policy, Permissions-Policy on https://cgi.com/.

CG-H3

Massive third-party script allowlist

HIGH
The problem

https://cgi.com/ response headers — CSP permits analytics/ads/chat/marketing hosts at scale (HubSpot, ZoomInfo, Crazy Egg, Facebook, Bing, New Relic, Leadoo, Siteimprove, …). last_reprobe=2026-07-31.

Weak browser security headers increase impact of XSS/clickjacking/MITM.

What an attacker or hacker agent can do
  • Human: Public headers surface on cgi.com expands attacker recon.
  • Agent (LLM+tools): Keep surface in continuous recon.
  • If invasive/destructive: Escalate if second bug/credential. L3 public-max is read-only.
Who owns it

CGI Inc. — security / product / infrastructure (confirm internal owner)

PoC (proof — does not fix)

Read-only. Does not change the system.

curl -sSI 'https://cgi.com/' | egrep -i 'strict-transport|content-security|x-frame|x-content'
How to fix

Add/tighten HSTS, CSP, XFO, nosniff, Referrer-Policy, Permissions-Policy on https://cgi.com/.

How to validate the fix

Add/tighten HSTS, CSP, XFO, nosniff, Referrer-Policy, Permissions-Policy on https://cgi.com/.

CG-M1

Large Drupal/JS surface

MEDIUM
The problem

Public JS bundles under /sites/default/files/js/; multi-MB sitemap. (surface: https://cgi.com/). last_reprobe=2026-07-31.

Public brand surface on cgi.com expands recon until closed.

What an attacker or hacker agent can do
  • Human: Public edge surface on cgi.com expands attacker recon.
  • Agent (LLM+tools): Keep surface in continuous recon.
  • If invasive/destructive: Escalate if second bug/credential. L3 public-max is read-only.
Who owns it

CGI Inc. — security / product / infrastructure (confirm internal owner)

PoC (proof — does not fix)

Read-only. Does not change the system.

curl -sSI 'https://cgi.com/'
How to fix

Remove unnecessary public hosts on cgi.com DNS or enforce strong auth/WAF; re-test from internet.

How to validate the fix

Remove unnecessary public hosts on cgi.com DNS or enforce strong auth/WAF; re-test from internet.

CG-M2

aide-dev.cgi.com public IIS 200

MEDIUM
The problem

CSP permits analytics/ads/chat/marketing hosts at scale (HubSpot, ZoomInfo, Crazy Egg, Facebook, Bing, New Relic, Leadoo, Siteimprove, …). last_reprobe=2026-07-31.

Weak browser security headers increase impact of XSS/clickjacking/MITM.

What an attacker or hacker agent can do
  • Human: Public headers surface on cgi.com expands attacker recon.
  • Agent (LLM+tools): Keep surface in continuous recon.
  • If invasive/destructive: Escalate if second bug/credential. L3 public-max is read-only.
Who owns it

CGI Inc. — security / product / infrastructure (confirm internal owner)

PoC (proof — does not fix)

Read-only. Does not change the system.

curl -sSI 'https://aide-dev.cgi.com' | egrep -i 'strict-transport|content-security|x-frame|x-content'
How to fix

Add/tighten HSTS, CSP, XFO, nosniff, Referrer-Policy, Permissions-Policy on https://cgi.com/.

How to validate the fix

Add/tighten HSTS, CSP, XFO, nosniff, Referrer-Policy, Permissions-Policy on https://cgi.com/.

CG-M3

accel360 / appstore / brand edges public

MEDIUM
The problem

Production script-src includes 'unsafe-inline' and 'unsafe-eval' plus a very large host allowlist. last_reprobe=2026-07-31.

Weak browser security headers increase impact of XSS/clickjacking/MITM.

What an attacker or hacker agent can do
  • Human: Public headers surface on cgi.com expands attacker recon.
  • Agent (LLM+tools): Keep surface in continuous recon.
  • If invasive/destructive: Escalate if second bug/credential. L3 public-max is read-only.
Who owns it

CGI Inc. — security / product / infrastructure (confirm internal owner)

PoC (proof — does not fix)

Read-only. Does not change the system.

curl -sSI 'https://cgi.com/' | egrep -i 'strict-transport|content-security|x-frame|x-content'
How to fix

Add/tighten HSTS, CSP, XFO, nosniff, Referrer-Policy, Permissions-Policy on https://cgi.com/.

How to validate the fix

Add/tighten HSTS, CSP, XFO, nosniff, Referrer-Policy, Permissions-Policy on https://cgi.com/.

Evidence annexes (AWP + Core)

Real pack from L3 max retest 2026-07-31 (public-max, non-destructive). Each file has a caption. Links open the artifact.

1) awp-receipt.json — AWP crypto

What it is: signed Agent Witness Protocol receipt (offline-verifiable).
What it proves: integrity of our test/engagement log.
What it is NOT: not alone proof of the bug on your host (that is each finding’s PoC).
→ download awp-receipt.json

2) leaves.jsonl — engagement diary

What it is: append-only leaves (scope, probes seal, catalog, quality, seal).
What it proves: timeline of this retest run.
→ download leaves.jsonl · leaf-chain.txt (PASS)

3) retest-raw.json — HTTP probes

What it is: status/headers/samples from public-max probes.
What it proves: what the internet saw during the run.
→ download retest-raw.json

4) agentic-walk.json — surface map

What it is: multi-page walk summary (URLs, sizes, forms).
→ download agentic-walk.json · core-export.json

5) folder

→ open full evidence folder · engagement eng_cgi_l3_retest_1361 · witness_mode local_awp_dev (DEV_LAB keys — not production ceremony).

How to use / verify AWP (open-source)

Client steps
  1. Download awp-receipt.json from the annex (or clone the evidence folder).
  2. On any machine with Node 18+:
# from the evidence folder
npx agent-witness-protocol verify awp-receipt.json

# or with local package
node /path/to/agent-witness-protocol/bin/awp.js verify awp-receipt.json

Honest PASS meaning: the receipt’s signature, Merkle inclusion, and checkpoint check out — the test log is intact.
PASS does NOT mean: your production bug is fixed, or that PayBotFin live network witness was used (this pack is local_awp_dev / DEV_LAB unless noted).

PayBotFin Witness: live network emit is optional and engagement-scoped. If not listed above as network_emit, do not claim live PayBot verification for this run.

Remediation order for the team

  1. Close HIGH public exposures first (auth / VPN / DNS / policy).
  2. Fix MEDIUM headers, mail policy, and non-prod exposure.
  3. Re-test every PoC from the public internet until closed.
  4. Archive AWP/evidence pack for audit and retest baseline.