| 1 |
Progeny Package Signing Policy |
| 2 |
|
| 3 |
John Goerzen <[email protected]> |
| 4 |
|
| 5 |
Draft 2; February, 2001 |
| 6 |
|
| 7 |
1 Introduction & Background |
| 8 |
|
| 9 |
One weakness commonly associated with installing software off the |
| 10 |
Internet – or even a CD-ROM – is verification of authenticity. |
| 11 |
Can a system administrator really be sure that the packages he |
| 12 |
thinks he is downloading from a Debian or Progeny mirror are |
| 13 |
really official packages? Or that a CD-ROM contains the original |
| 14 |
packages? These days, it is not impossibly difficult to intercept |
| 15 |
network traffic and insert trojan packages in place of real ones, |
| 16 |
or to prepare a fake CD-ROM with a realistic-looking label. To |
| 17 |
thwart this, the ability to add GPG signatures to Debian packages |
| 18 |
is being implemented. These signatures add a cryptographic |
| 19 |
guarantee that the package is authentic in a particular way. |
| 20 |
|
| 21 |
Each package may carry multiple signatures, and each signature |
| 22 |
may convey unique information. For instance, one signature might |
| 23 |
confirm that a legitimate Progeny developer uploaded the package; |
| 24 |
another, that Progeny Quality Assurance has processed the |
| 25 |
package; and a final one might confirm that a package is part of |
| 26 |
a particular release. |
| 27 |
|
| 28 |
On each customer's computer is a file containing rules for |
| 29 |
authentication of these GPG signatures. The rules file describes |
| 30 |
which signatures are required, which are optional, and what keys |
| 31 |
constitute a valid signature for each signature type. Therefore, |
| 32 |
Progeny can ship rules that might say something like this: “For |
| 33 |
each package to be installed, require the Progeny corporate |
| 34 |
signature, a signature from the Quality Assurance team, and |
| 35 |
either a signature from a Progeny developer that uploaded a |
| 36 |
Progeny package or an automated signature indicating that the |
| 37 |
package came from Debian. |
| 38 |
|
| 39 |
1.1 Goals |
| 40 |
|
| 41 |
There are several things that we would like to accomplish with |
| 42 |
this system. Here is a list of many of them, an explanation of |
| 43 |
how we accomplish it, and why it is valuable. |
| 44 |
|
| 45 |
1.1.1 Progeny Developer Verification |
| 46 |
|
| 47 |
The first thing to check on a package is to verify that it was |
| 48 |
built by an authorized Progeny developer. To this end, we will |
| 49 |
have a maintainer (“maint”) signature applied to each package at |
| 50 |
upload time by the person that built it. This key provides |
| 51 |
several benefits: |
| 52 |
|
| 53 |
• It provides a guarantee that the package was not modified by |
| 54 |
anyone else within Progeny. No malicious Progeny employee could |
| 55 |
modify the package on its way. |
| 56 |
|
| 57 |
• Customers can know that a package was unmodified from the time |
| 58 |
it was built. |
| 59 |
|
| 60 |
Additionally, if Progeny ever discovers that a particular |
| 61 |
developer was malicious or dangerously incompetent in the |
| 62 |
building of packages, the packages built by that person can be |
| 63 |
easily identified and can be rendered uninstallable. |
| 64 |
|
| 65 |
1.1.2 Quality Assurance Verification |
| 66 |
|
| 67 |
Another useful point to check on a package is whether or not it |
| 68 |
has passed through Progeny's Quality Assurance department. By |
| 69 |
doing so, we can guarantee that no package bypassed the QA |
| 70 |
department (either by accident or maliciously), and introduce an |
| 71 |
additional level of verification. By having this second signature |
| 72 |
present on the package, we reduce the damage done if the key for |
| 73 |
an individual maintainer were to become compromised in some way. |
| 74 |
Because we have a maintainer signature already, we have already |
| 75 |
prevented any malicious QA employee from modifying the package, |
| 76 |
but if Progeny does at some time find cause to suspect a QA |
| 77 |
person of handling packages in a dangerous way, the key can be |
| 78 |
removed and the packages rendered uninstallable in the same |
| 79 |
manner as with the Debian maintainer signature. |
| 80 |
|
| 81 |
1.1.3 General Malicious Employee Protection |
| 82 |
|
| 83 |
Because Progeny's business is so highly dependent upon secure and |
| 84 |
stable software, if a malicious employee were to be ever hired, |
| 85 |
serious damage could be done. Package signatures and |
| 86 |
compartmentalization can be used together to provide an effective |
| 87 |
weapon to significantly lessen the potential harm that could be |
| 88 |
caused by such a person. In general, the worst harm that could be |
| 89 |
inflicted upon Progeny by a malicious employee would be the |
| 90 |
insertion of trojan or otherwise malicious code into the packages |
| 91 |
that we distribute to customers and use ourselves. The mechanisms |
| 92 |
outlined above can help to combat this threat in several ways. |
| 93 |
|
| 94 |
• A malicious system administrator with full access to the |
| 95 |
machines on which the archive is served from would ordinarily |
| 96 |
be able to modify the packages on those machines and insert |
| 97 |
trojan code virtually undetected. Since the packages had |
| 98 |
already gone through QA, and probably were already installed on |
| 99 |
Progeny systems, the change would probably be noticed by |
| 100 |
customers first. With this system, the maintainer signature is |
| 101 |
applied before the package is even uploaded to the archive in |
| 102 |
the first place, so any such change would be immediately |
| 103 |
noticeable. |
| 104 |
|
| 105 |
• A malicious developer, often also with full access to the |
| 106 |
archive, could bypass QA to get malicious code into the |
| 107 |
archive. With this system, no package would be installable |
| 108 |
unless it bears a signature from a valid QA employee. |
| 109 |
|
| 110 |
• A malicious QA employee could modify the package after it has |
| 111 |
been received from the developers but before it is placed onto |
| 112 |
the archive. With this proposal, such a change would be |
| 113 |
immediately noticed because the maintainer signature check |
| 114 |
would fail. |
| 115 |
|
| 116 |
1.1.4 General Malicious Outsider Protection |
| 117 |
|
| 118 |
Another category of problems that we have to worry about concern |
| 119 |
malicious outsiders. There are several ways that such a person |
| 120 |
could cause Progeny customers to unknowingly install and run code |
| 121 |
that poses a security risk to them. We can deal with most of |
| 122 |
them. |
| 123 |
|
| 124 |
• Someone could intercept the traffic between a customer's |
| 125 |
computer and the Progeny archive, inserting data containing |
| 126 |
trojan code without the customer's knowledge. This is already |
| 127 |
possible and has been done. By placing signatures on the |
| 128 |
packages, if they are modified any way during download, the |
| 129 |
customer would immediately be aware. |
| 130 |
|
| 131 |
• By “social engineering”, that is, convincing someone through |
| 132 |
non-technical means to install a package, a malicious person |
| 133 |
could convince one of our customers to install a package that |
| 134 |
contains a security problem. However, with signature |
| 135 |
verification, such a package would fail to install. |
| 136 |
|
| 137 |
• An outsider could replace the data in a current package with |
| 138 |
that from a previous version of a package, perhaps one with |
| 139 |
known security vulnerabilities. With our archive and release |
| 140 |
signatures, we can dramatically reduce the potential pool of |
| 141 |
packages from which such a person could draw a signed Progeny |
| 142 |
package containing a known security problem. |
| 143 |
|
| 144 |
• A cracker or someone else that gains unauthorized access to |
| 145 |
Progeny archives could modify the .debs that customers download |
| 146 |
in-place, giving them code that could pose a security risk. |
| 147 |
With the mechanisms in this proposal, such a modification would |
| 148 |
cause multiple signatures to fail, providing an instant way to |
| 149 |
know that the package had been modified. |
| 150 |
|
| 151 |
1.2 Unaddressed Issues |
| 152 |
|
| 153 |
While this proposal can deal with a great many potential problems |
| 154 |
that could occur, there are still several problems that cannot be |
| 155 |
dealt with solely within the scope of this system. Here are some |
| 156 |
of them that we should be aware of. |
| 157 |
|
| 158 |
• A cracker obtaining root maliciously on a customer's machine |
| 159 |
could bypass all of the checks that are in place here. Such a |
| 160 |
person could install fake versions of the verifier that could |
| 161 |
report false positives. In general, if root is compromised, we |
| 162 |
can no longer make any assumptions about the security of the |
| 163 |
customer's machine. |
| 164 |
|
| 165 |
• A cracker or internal person that has obtained root on a |
| 166 |
Progeny developer's machine could alter files such that |
| 167 |
packages built by that developer would contain trojan code that |
| 168 |
the developer may not be aware of. Hopefully, this would be |
| 169 |
detected during the QA process, but there is no guarantee of |
| 170 |
that. Such a cracker could also replace GPG with a program that |
| 171 |
could surreptitiously gather the developer's secret key and |
| 172 |
passphrase, causing a compromise of the developer's secret key. |
| 173 |
Again, once someone has obtained root, we can no longer make |
| 174 |
any assumptions about the security of the machine. |
| 175 |
|
| 176 |
• We must have a secure way to get the initial keyring and |
| 177 |
software to the customer. This is a question of how far we go |
| 178 |
to make the guarantee. Some options include placing the |
| 179 |
software behind a HTTPS site that bears a signature from a |
| 180 |
well-known certifying authority. However, if someone's web |
| 181 |
browser has already been modified by a malicious person, even |
| 182 |
this would not guarantee success. Another option is to |
| 183 |
distribute the initial software and keys via CD-ROM shipped |
| 184 |
directly from our office, though once again, a malicious person |
| 185 |
could prepare a fake CD-ROM bearing a realistic-looking label |
| 186 |
and ship it to a customer. These are questions that we have to |
| 187 |
deal with outside the framework of the package signing software |
| 188 |
itself, and basically rely upon striking an acceptable balance |
| 189 |
between paranoia and ease of use. |
| 190 |
|
| 191 |
2 Criteria For Verification |
| 192 |
|
| 193 |
The verification process is that which is performed on a |
| 194 |
customer's machine when a package is to be installed. |
| 195 |
Verification software has configurable criteria to indicate what |
| 196 |
signatures are necessary to consider a package to be valid. While |
| 197 |
the verification policy is up to each individual system |
| 198 |
administrator, we must supply a default. A package from Progeny |
| 199 |
will consider to pass the tests if all of these criteria are met: |
| 200 |
|
| 201 |
• The package must bear an “origin” signature made by a |
| 202 |
designated origin key. |
| 203 |
|
| 204 |
• The package must bear a quality assurance (“qa”) signature made |
| 205 |
by a member of the Progeny QA team. |
| 206 |
|
| 207 |
• The package must have a maintainer (“maint”) signature from |
| 208 |
either: |
| 209 |
|
| 210 |
|
| 211 |
– A member of the Progeny development team; or |
| 212 |
|
| 213 |
– An automated signature applied to packages imported |
| 214 |
unmodified from Debian. |
| 215 |
|
| 216 |
Possible modifications that we could consider making would be to |
| 217 |
change some of the logic from “and” to “or”, thus permitting |
| 218 |
installation even if one signature is missing or somehow invalid. |
| 219 |
This could make dealing with situations involving lost keys, |
| 220 |
changing employees, and upgrades easier, but comes at the expense |
| 221 |
of opening up some potential malicious employee attacks. |
| 222 |
|
| 223 |
3 Procedures For Signing |
| 224 |
|
| 225 |
Keys can be added to packages during the development cycle. At |
| 226 |
each step, we should ensure that all existing keys on the package |
| 227 |
are valid. The first signature on a package will be added by |
| 228 |
either the Progeny developer responsible for it or the system |
| 229 |
that is installing the package from Debian. The next signature |
| 230 |
will be added by the person in QA that approves the package. |
| 231 |
After that, the origin signature should be added, and finally, |
| 232 |
when the package is placed into the archive, an archive signature |
| 233 |
should be added. All signatures except automated ones are to be |
| 234 |
made by personal keys. Progeny will maintain keyrings of valid |
| 235 |
developer and QA keys. Below, the meaning and procedures for each |
| 236 |
Progeny signature are described. |
| 237 |
|
| 238 |
3.1 Maintainer Signature |
| 239 |
|
| 240 |
The maintainer signature is named “maint”. It is to be applied to |
| 241 |
any package uploaded into the archive by a Progeny developer. The |
| 242 |
maintainer signature is made with that developer's own GPG key. |
| 243 |
Its purpose is to let the customer verify that the package was |
| 244 |
indeed worked on by an authorized Progeny employee. |
| 245 |
|
| 246 |
Many packages (a majority, in fact) appear in the Progeny |
| 247 |
distribution directly from Debian with no modification or |
| 248 |
intervention from Progeny employees. There is no real Progeny |
| 249 |
developer or uploader in charge of these packages. For these |
| 250 |
situations, a special key should be generated. A signature from |
| 251 |
that key should be automatically added to packages that come |
| 252 |
through from Debian. The username and description for that key |
| 253 |
should make clear that it does not originate from a particular |
| 254 |
Progeny developer. |
| 255 |
|
| 256 |
3.2 Quality Assurance |
| 257 |
|
| 258 |
As packages move into the distribution, we anticipate that they |
| 259 |
will be screened by QA. To that end, a “qa” signature will be |
| 260 |
made for the package. It will be made using the GPG key of the |
| 261 |
specific QA employee that approves the package. Any package |
| 262 |
bearing a valid QA signature can thus be assumed to have passed |
| 263 |
their testing for the distribution. There should be no case in |
| 264 |
which this key is added automatically. An exception may need to |
| 265 |
be made for initial bootstrapping; that is, dealing with the |
| 266 |
thousands of packages already in the distribution prior to QA |
| 267 |
procedures. |
| 268 |
|
| 269 |
3.3 Origin |
| 270 |
|
| 271 |
The only key that must be present in the debsigs system is “ |
| 272 |
origin”. This key certifies the originator of the package (in |
| 273 |
this case, Progeny Linux Systems) and lets the signature verifier |
| 274 |
load the appropriate ruleset for the verification of signatures. |
| 275 |
This key will probably be added automatically, using our |
| 276 |
corporate key or some designated origin key that bears its |
| 277 |
signature. |
| 278 |
|
| 279 |
4 Safeguarding Keys |
| 280 |
|
| 281 |
The task of safeguarding individual keys is already well-known, |
| 282 |
understood, and easily manageable. However, safeguarding keys |
| 283 |
that are used by many people or to apply automated signatures |
| 284 |
gets more difficult. In some cases, the security of these is not |
| 285 |
as important – for instance, an archive key will not inflict much |
| 286 |
damange if compromised. However, in other cases, the security is |
| 287 |
still quite important. |
| 288 |
|
| 289 |
Safeguarding of individual keys is a well-known problem with |
| 290 |
well-documented answers. To provide a comprehensive system, there |
| 291 |
are also other cases that we need to consider, which are spelled |
| 292 |
out below. |
| 293 |
|
| 294 |
5 Special Cases & Procedures |
| 295 |
|
| 296 |
While the policy laid out here will work smoothly for day-to-day |
| 297 |
operation, things can go wrong or get confused and planning for |
| 298 |
them is a must. In this section, I describe some special cases |
| 299 |
that can face us, what they mean for the package signing system, |
| 300 |
and how we can prepare for them ahead of time and deal with them |
| 301 |
when they occur. |
| 302 |
|
| 303 |
5.1 Employee Leaves |
| 304 |
|
| 305 |
If a Progeny employee leaves the company for some reason, and |
| 306 |
this person was someone whose key was used to create signatures |
| 307 |
for packages that are used by customers, then Progeny will need |
| 308 |
to decide, on a case-by-case basis, several things: |
| 309 |
|
| 310 |
• Do we have any reason to believe that the packages in the |
| 311 |
archive now bearing this person's signature are in any way |
| 312 |
damaged or tampered with? |
| 313 |
|
| 314 |
• Do we have any reason to believe that this person may, in the |
| 315 |
future, intend to inflict harm on Progeny through the use of |
| 316 |
the key used to apply those signatures? |
| 317 |
|
| 318 |
If the answer to the first question is “yes”, then we would |
| 319 |
immediately need to remove the key of this employee from all |
| 320 |
keyrings and rebuild the effected packages. If the answer to the |
| 321 |
second question is “yes”, then we would need to immediately |
| 322 |
remove this person's key from all internal keyrings (such that |
| 323 |
the maintainer signature check made by QA would fail if somehow |
| 324 |
the person uploads a package after leaving the company), but |
| 325 |
removing the key from the external keyrings is not a priority and |
| 326 |
can be done at a more leisurely pace. If both answers are “no”, |
| 327 |
then Progeny should, as a precaution, rebuild packages signed by |
| 328 |
that user eventually, but there is no need to do so before the |
| 329 |
next release. |
| 330 |
|
| 331 |
5.2 People Join Progeny |
| 332 |
|
| 333 |
When a new person is hired for a development or QA position, this |
| 334 |
person will need to generate a personal key and have it inserted |
| 335 |
onto the keyring. Customers will not be able to verify a |
| 336 |
signature made by this new person until they have upgraded their |
| 337 |
Progeny keyring package. Once that is done, packages will be able |
| 338 |
to verify with no difficulty. We may need to ensure that the new |
| 339 |
keyring package gets installed early in any upgrade procedure, |
| 340 |
but beyond that, there is little that we must do specially in |
| 341 |
this case. Note, though, that an existing employee must be the |
| 342 |
one to sign the new keyring package. |
| 343 |
|
| 344 |
5.3 Many Employees Leave |
| 345 |
|
| 346 |
In the event that enough employees from a particular department |
| 347 |
leave the company such that there no longer exists anyone with a |
| 348 |
key that can make a valid signature for a particular category, we |
| 349 |
need to have a way to allow customers to continue receiving |
| 350 |
updates in a secure fashion. For this purpose, I suggest that |
| 351 |
Progeny create a generic key for both the QA and maintainer |
| 352 |
categories, store it in some physically secure area (such as a |
| 353 |
safe or deposit box) on removable media in a fashion that only |
| 354 |
authorized company officers could get the key in an emergency. |
| 355 |
The public key would be inserted on the distributed keyring |
| 356 |
files, but would never be used unless absolutely necessary. |
| 357 |
|
| 358 |
5.4 Key Compromises |
| 359 |
|
| 360 |
If a key becomes compromised (that is, an intruder obtains the |
| 361 |
secret key and passphrase, enabling unauthorized signatures), |
| 362 |
then we must take steps to rectify the problem immediately. |
| 363 |
Depending on the type of key that was compromised and the |
| 364 |
circumstances, we might opt to immediately remove the compromised |
| 365 |
key from all keyrings and rebuild any effected packages. In all |
| 366 |
cases, the victim of the compromise should issue a revocation |
| 367 |
certificate, discard the old key, and generate a new one for |
| 368 |
future use. |
| 369 |
|
| 370 |
5.5 Compartmentalization |
| 371 |
|
| 372 |
One of the benefits of having multiple signatures on a package is |
| 373 |
to reduce the damage that can be inflicted if one of those keys |
| 374 |
becomes untrusted in some fashion. For this reason, we should |
| 375 |
make sure that we do not have overlap between those that make |
| 376 |
signatures for different parts of the operation. Specifically, |
| 377 |
nobody should be both on the maintainer and the QA keyring. |
| 378 |
|
| 379 |
6 Future Options |
| 380 |
|
| 381 |
In the course of preparing this policy, some ideas were discussed |
| 382 |
but discarded for the moment. This section will explain them as |
| 383 |
possible options for future use. |
| 384 |
|
| 385 |
6.1 Goal: Release & Dating of Packages |
| 386 |
|
| 387 |
One potential problem that could arise for a customer is that a |
| 388 |
malicious person could present an old package as if it were |
| 389 |
current. If this old package contains problems, such as known |
| 390 |
security vulnerabilities, doing so could be an avenue for attack. |
| 391 |
Preventing this problem entirely is not really feasible; however, |
| 392 |
we can significantly lessen the potential vulnerability by |
| 393 |
providing a limited “window” of time for which we consider a |
| 394 |
particular package valid. |
| 395 |
|
| 396 |
For this situation, we must consider CD-ROM and online |
| 397 |
distribution separately. For online distribution, we can apply an |
| 398 |
“archive” signature, automatically, to each package in the |
| 399 |
archive. Because all signatures contain a timestamp indicating |
| 400 |
when they were made, the signature verifier on the customer's |
| 401 |
machine can ensure that the signature was placed on the package |
| 402 |
within a certain timeframe – for example, two weeks. On our side, |
| 403 |
we would automatically re-sign all packages with the archive key |
| 404 |
periodically – in this example, perhaps once a week. In this way, |
| 405 |
a customer installing a package knows that it is no more than two |
| 406 |
weeks out of date. Packages distributed via CD-ROM will either |
| 407 |
not have the archive signature or will have it ignored. |
| 408 |
|
| 409 |
This solution won't work for a CD-ROM because they are frequently |
| 410 |
used more than two weeks from the release and because there is no |
| 411 |
way to propagate signature updates to people using them. |
| 412 |
Therefore, the best we can do is guarantee the customer that a |
| 413 |
given package is indeed part of the release that is expected – |
| 414 |
that is, if Progeny 2.1 is current, that a given package was |
| 415 |
released as part of Progeny 2.1 and does not date back to version |
| 416 |
1.1. Another benefit of the release signature is that it can be |
| 417 |
virtually uncompromisable. The release signature can be applied |
| 418 |
to the packages at release time, and then the key behind it can |
| 419 |
be destroyed. However, as a practical matter, we will need to |
| 420 |
publish at least one future release key in the current keyring so |
| 421 |
that people upgrading will be able to do so in a secure fashion. |
| 422 |
|
| 423 |
6.1.1 Archive |
| 424 |
|
| 425 |
The “archive” signature is to be automatically applied by |
| 426 |
dinstall when a package is installed into Progeny's |
| 427 |
publically-accessible archive, regardless of the particular area |
| 428 |
that it is installed into. A valid signature can be used to |
| 429 |
assure that a package was indeed posted by Progeny on a specific |
| 430 |
date. |
| 431 |
|
| 432 |
6.1.2 Release |
| 433 |
|
| 434 |
The release key should be kept in a secure, safe place accessible |
| 435 |
only by authorized company officers prior to use. After it has |
| 436 |
been used, it should be permanently destroyed. Release keys will |
| 437 |
need to be generated at least one release ahead of use to provide |
| 438 |
a suitable upgrade path for users. |
| 439 |
|